Full Reading Guide

Agentic AI for Utility Operators — integrated power-system engineering tutorial. Evidence reviewed 17 October 2026. The full reading guide is separate from the timed main talk. Corporate issuers are listed when no individual author is recorded. Original figures, page locations and image-source credits are included. Attribution alone is not a blanket reuse license.

[B01] IEEE Communications Society: official resources and network presentation template

Authors / issuer
IEEE Communications Society
Publisher / venue
IEEE Communications Society
Date
Reviewed 17 October 2026
Theme assets
comsoc-lockup.png and image7.png; independent educational adaptation
Full source URL
https://www.comsoc.org/about/brand-experience/resources

[DEF03] Building effective agents

Authors / issuer
Erik S., Barry Zhang
Publisher / venue
Anthropic Engineering
Date
2024-12-19
Type
Official practitioner architecture account by the system maintainers
Location
What are agents?; Agents; Workflow: Routing; Workflow: Orchestrator-workers; Workflow: Evaluator-optimizer
Version note
The live page now warns that tooling has changed since the original publication. Use its conceptual distinction, not its present SDK list as a claim about the 2024 implementation.
Full source URL
https://www.anthropic.com/engineering/building-effective-agents

[DEF02] Intelligent agents: theory and practice

Authors / issuer
Michael Wooldridge, Nicholas R. Jennings
Publisher / venue
The Knowledge Engineering Review, 10(2), pp. 115–152
Date
1995
Type
Primary theoretical/architectural research review by the authors
Location
Section 1.1, What is an Agent?; Section 1.1.1, A Weak Notion of Agency
Version note
The author-hosted HTML is an earlier draft stamped November 1994. Use the final 1995 PDF for the bibliographic citation; the HTML makes the definition sections searchable. The final PDF is image-based in the web extractor, so section locators are safer than an unverified PDF page number.
Full source URL
https://doi.org/10.1017/S0269888900008122
Full paper PDF
https://www.cs.ox.ac.uk/people/michael.wooldridge/pubs/ker95.pdf

[NV01] Building for the Rising Complexity of Agentic Systems with Extreme Co-Design

Authors / issuer
Eduardo Alvarez, Benjamin Klieger, Graham Steele
Publisher / venue
NVIDIA Technical Blog
Date
May 5, 2026
Type
Official vendor architectural explanation; conceptual figure
Location
Transition to agents from chatbots, Figure 1
Full source URL
https://developer.nvidia.com/blog/building-for-the-rising-complexity-of-agentic-systems-with-extreme-co-design/

[GEN01] Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile

Authors / issuer
National Institute of Standards and Technology
Publisher / venue
NIST AI 600-1
Date
July 2024
Type
Official primary definition and risk-management profile
Location
Introduction, printed p.1, footnote 1; retrieved 18 October 2026
Full source URL
https://doi.org/10.6028/NIST.AI.600-1
Full paper PDF
https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf

[DEF01] Artificial Intelligence: A Modern Approach, Chapter 2: Intelligent Agents

Authors / issuer
Stuart J. Russell, Peter Norvig
Publisher / venue
Prentice Hall, second edition, Chapter 2, pp. 32–58
Date
2003
Type
Canonical authored textbook; primary exposition of its agent framework
Location
Section 2.1, printed pp. 32-34 / PDF pp. 1-3; Figure 2.1, printed p. 33 / PDF p. 2; Section 2.2, printed pp. 35-37 / PDF pp. 4-6
Version note
Despite its current /4th-ed/ URL, the verified sample chapter is the older chapter with printed pages 32-58 and 2002 directory timestamp. Do not cite this PDF as fourth-edition text. The official second-edition page identifies the edition and Prentice Hall; the final book citation is 2003.
Full source URL
https://aima.cs.berkeley.edu/2nd-ed/
Full paper PDF
https://aima.cs.berkeley.edu/4th-ed/pdfs/newchap02.pdf

[DEF06] Tools

Authors / issuer
LangChain maintainers
Publisher / venue
Docs by LangChain, Python open-source documentation
Date
Accessed 17 October 2026
Type
Official tool-interface documentation
Full source URL
https://docs.langchain.com/oss/python/langchain/tools

[NV02] Introduction to LLM Agents

Authors / issuer
Tanay Varshney
Publisher / venue
NVIDIA Technical Blog
Date
November 30, 2023
Type
Official vendor architectural explanation; conceptual figure
Location
What is an AI agent? Figure 1: General components of an agent
Full source URL
https://developer.nvidia.com/blog/introduction-to-llm-agents/

[VID03] Agentic AI: Workflows vs. agents

Authors / issuer
Aja Hammerly, Jason Davenport
Publisher / venue
Google Cloud Tech / YouTube
Date
19 March 2025
Evidence type
Official conceptual explainer; general software examples
Full source URL
https://www.youtube.com/watch?v=Qd6anWv0mv0

[LG03] LangGraph Interrupts

Authors / issuer
LangChain documentation maintainers
Publisher / venue
Official LangChain documentation
Date
Accessed 2026-10-17; live documentation
Evidence type
Official Python runtime documentation
Full source URL
https://docs.langchain.com/oss/python/langgraph/interrupts

[DEF04] LangGraph overview

Authors / issuer
LangChain maintainers
Publisher / venue
Docs by LangChain, Python open-source documentation
Date
Undated live documentation; accessed 2026-10-17
Type
Official framework and tool-interface documentation; two pages in one documentation collection
Location
LangGraph overview: opening paragraphs and product-role list; Tools: opening definition and Basic tool definition
Full source URL
https://docs.langchain.com/oss/python/langgraph/overview

[PA02] PowerAgent — Open-Source Agentic AI for Power Grids

Authors / issuer
Power and AI Initiative / PowerAgent community, Harvard SEAS
Publisher / venue
Official Harvard project website
Date
Live project page reviewed 17 October 2026
Evidence type
Primary project scope, maintenance and demonstration descriptions
Full source URL
https://poweragent.seas.harvard.edu/

[PD01] PowerDAG: Supervisory Agentic AI System for Automating Distribution Grid Analysis

Authors / issuer
Emmanuel O. Badmus, Amritanshu Pandey
Publisher / venue
arXiv preprint, version 4
Date
5 July 2026 (v4); first submitted 18 March 2026
Evidence type
Full primary research paper; v4 inspected
Full source URL
https://arxiv.org/pdf/2603.17418v4

[N01] Field Notes • HyperQ deployment in PJM’s Cycle 1

Authors / issuer
Tapestry
Publisher / venue
Tapestry
Date
2026-06-11; deployment entry 2026-04-29
Type
supplier primary deployment/results report
Full source URL
https://www.tapestryenergy.com/en/projects/field-notes-hyperq-deployment-in-pjm-s-cycle-1

[N02] How agentic AI can help grid operators speed up interconnection

Authors / issuer
Tapestry
Publisher / venue
Tapestry
Date
2025-12-08 (linked chronological entry in N01); live article reviewed 2026-10-17
Type
supplier primary technical white paper
Full source URL
https://www.tapestryenergy.com/en/projects/how-agentic-ai-can-help-grid-operators-speed-up-interconnection

[N06] Successful Interconnection Reforms, Other PJM Initiatives Seek To Maximize Electricity Supplies

Authors / issuer
PJM Interconnection
Publisher / venue
PJM Interconnection
Date
2026-03-16
Type
grid-operator primary progress report
Full source URL
https://insidelines.pjm.com/connected/

[C01] Exploring AI to support smarter grid management

Authors / issuer
California ISO
Publisher / venue
California ISO
Date
2025-07-28
Type
customer primary source
Full source URL
https://www.caiso.com/about/news/energy-matters-blog/exploring-ai-to-support-smarter-grid-management

[C02] Generative and Agentic AI in Grid Operations: Field Lessons from OATI, CAISO, PNNL, and EPRI

Authors / issuer
OATI
Publisher / venue
OATI
Date
2026 (exact publication day not displayed)
Type
vendor primary field report
Full source URL
https://www.oati.com/ai-for-energy/

[I06] California ISO control room, image published by OATI: California ISO control room with operators seated at curved desks facing a long wall of grid displays

Authors / issuer
OATI (California ISO control-room photograph)
Publisher / venue
OATI (California ISO control-room photograph)
Date
Image accessed 17 October 2026
Full source URL
https://www.oati.com/ai-genie/

[U02] AWS Launches Agentic Grid Planning Program to Accelerate Interconnection Studies

Authors / issuer
Amazon Web Services
Publisher / venue
Amazon Web Services
Date
2026-09-17
Type
vendor primary announcement with named Duke Energy collaboration and testimony
Full source URL
https://press.aboutamazon.com/aws/2026/9/aws-launches-agentic-grid-planning-program-to-accelerate-interconnection-studies

[U03] Agentic Grid Planning on AWS

Authors / issuer
Amazon Web Services
Publisher / venue
Amazon Web Services
Date
Undated; reviewed 17 October 2026
Type
vendor primary program and workflow description
Full source URL
https://aws.amazon.com/energy-utilities/agentic-grid-planning/

[I03] Duke Energy, illumination, August 2, 2022: Duke Energy grid operator Howard Greene seated before many control-room displays

Authors / issuer
Duke Energy / illumination
Publisher / venue
Duke Energy / illumination
Date
Image accessed 17 October 2026
Full source URL
https://illumination.duke-energy.com/articles/meet-the-team-that-monitors-the-grid-and-keeps-the-lights-on

[N03] Siemens’ Gridscale X redefines system operations and agentic transmission planning

Authors / issuer
Siemens
Publisher / venue
Siemens
Date
2026-05-05
Type
vendor primary product announcement
Full source URL
https://press.siemens.com/global/en/pressrelease/siemens-gridscale-x-redefines-system-operations-and-agentic-transmission-planning

[N04] National Grid Full Year Results 2025/26: Video Webcast Transcript

Authors / issuer
National Grid
Publisher / venue
National Grid
Date
2026-05-14
Type
utility primary investor-results transcript
Full source URL
https://www.nationalgrid.com/document/577036/download

[N09] On demand: Trust-driven medallion architecture: How to make your CDEs ready for AI

Authors / issuer
Session speakers: Saptaparna Gupta (NYPA), Oleksandra Bovkun (Databricks), Eliza Pochinova (Ataccama)
Publisher / venue
Ataccama
Date
2026-10-14 (on-demand page); event page currently lists 2026-10-13
Type
Original vendor-hosted technical session report with named utility participant
Full source URL
https://www.ataccama.com/webinar/on-demand-trust-driven-medallion-architecture

[N10] Trust-driven medallion architecture: how to make your CDEs ready for AI

Authors / issuer
Saptaparna Gupta (NYPA), Oleksandra Bovkun (Databricks), Eliza Pochinova (Ataccama)
Publisher / venue
Ataccama
Date
2026-10-13, 11:00 AM EDT (date displayed on fetched event page)
Type
Original technical event listing with named utility speaker
Full source URL
https://www.ataccama.com/events/trust-driven-medallion-architecture

[N07] Distribution System Modeling and Automation Engineer (Hybrid Schedule)

Authors / issuer
Eversource Energy (employer)
Publisher / venue
Eversource Energy
Date
2026-10-05
Type
Utility primary engineering recruitment disclosure; capability-building evidence
Full source URL
https://jobs.eversource.com/job/east-berlin/distribution-system-modeling-and-automation-engineer-hybrid-schedule/7774/101576695504

[N11] Eversource Energy Collaborates with MathWorks to Incorporate Renewable Energy Sources into System Planning Process Using Probabilistic Load Flow Automation

Authors / issuer
MathWorks; named Eversource engineer John Kreso quoted
Publisher / venue
MathWorks
Date
2025-04-29
Type
Supplier primary engineering deployment report with named utility testimony; conventional automation
Full source URL
https://www.mathworks.com/company/newsroom/eversource-energy-collaborates-with-mathworks-to-incorporate-renewable-energy-sources-into-system-planning-process-using-probabilistic-load-flow-automation.html

[C07] GE Vernova Launches GridOS for Distribution, the Industry's First Unified Solution for Grid Orchestration

Authors / issuer
GE Vernova
Publisher / venue
GE Vernova
Date
2026-02-03
Type
vendor primary announcement with named utility statement
Full source URL
https://www.gevernova.com/news/press-releases/ge-vernova-launches-gridos-distribution-industrys

[I01] GE Vernova: GE Vernova GridOS promotional control-room composite with people and grid software displays

Authors / issuer
GE Vernova
Publisher / venue
GE Vernova
Date
Image accessed 17 October 2026
Full source URL
https://www.gevernova.com/software/products/gridos

[Q01] An Analysis of Grid Operators

Authors / issuer
Corey Fallon, Brett Jefferson, Tyler Ryan
Publisher / venue
Pacific Northwest National Laboratory
Date
2023-04-19
Type
Primary research presentation; 20-operator survey
Location
Methods PDF p.5; workload result PDF p.9. Exact survey question and task denominator are not supplied.
Full source URL
https://www.pnnl.gov/sites/default/files/media/file/PNNL_workshop_Presentation_Human_Factors_Virtual_Workshop_March_2023.pdf

[O01] What Happens in the Control Room

Authors / issuer
PJM Interconnection
Publisher / venue
PJM Interconnection
Date
2026-04-16
Type
Official operator fact sheet
Full source URL
https://www.pjm.com/-/media/DotCom/about-pjm/newsroom/fact-sheets/what-happens-in-the-control-room-fact-sheet.pdf

[I02] Duke Energy, illumination, August 2, 2022: Duke Energy employees in an Energy Control Center, with one pointing toward a display

Authors / issuer
Duke Energy / illumination
Publisher / venue
Duke Energy / illumination
Date
Image accessed 17 October 2026
Full source URL
https://illumination.duke-energy.com/articles/meet-the-team-that-monitors-the-grid-and-keeps-the-lights-on

[O05] Control center modernization is enhancing reliability

Authors / issuer
John Phipps
Publisher / venue
California ISO
Date
2025-09-10
Type
Official operator article and tool screenshot
Full source URL
https://www.caiso.com/about/news/energy-matters-blog/control-center-modernization-is-enhancing-reliability

[I09] GridOS Visual Intelligence product screens

Authors / issuer
GE Vernova
Publisher / venue
GE Vernova
Date
Accessed 17 October 2026
Full source URL
https://www.gevernova.com/software/products/gridos/visual-intelligence

[LG01] LangGraph Graph API overview

Authors / issuer
LangChain documentation maintainers
Publisher / venue
Official LangChain documentation
Date
Accessed 2026-10-17; live documentation
Evidence type
Official Python runtime documentation
Full source URL
https://docs.langchain.com/oss/python/langgraph/graph-api

[E01] X-GridAgent: An LLM-Powered Agentic AI System for Assisting Power Grid Analysis

Authors / issuer
Yihan (Logon) Wen, Xin Chen
Publisher / venue
arXiv preprint
Date
2025-12-23 (v1)
Evidence type
Original research; full v1 PDF
Full source URL
https://arxiv.org/pdf/2512.20789v1

[R01] Integrating Agentic Artificial Intelligence with High-Performance Computing for Grid Planning

Authors / issuer
Samim Konjicija, Slaven Peles
Publisher / venue
arXiv; EnergyCon 2026 proceedings forthcoming as described by authors
Date
3 September 2026; v1
Type
Research preprint; not a utility deployment
Location
Abstract and framework description: LLM-directed analyses through ExaGO; University of Sarajevo and Oak Ridge National Laboratory affiliations
Full source URL
https://arxiv.org/abs/2609.04544v1
Full paper PDF
https://arxiv.org/pdf/2609.04544v1

[L01] GridGuard current project: README and graph source snapshot

Authors / issuer
GridGuard project / local source snapshot
Publisher / venue
GridGuard project / local source snapshot
Date
Snapshot 17 October 2026; see research/GridGuard_source_snapshot.txt
Full source URL
research/GridGuard_source_snapshot.txt

[D01] DrAgent: An Agentic Approach to Fault Analysis in Power Grids Using Large Language Models

Authors / issuer
Barun Kumar Saha, Aarthi V, O. D. Naidu
Publisher / venue
2025 International Conference on Artificial Intelligence in Information and Communication (ICAIIC), pp. 938–945
Date
2025-02-18–21
Type
Peer-reviewed original power-system research; full conference-hosted PDF
Full source URL
https://www.manuscriptlink.com/society/kics/media?key=kics%2Fconference%2Ficaiic2025%2Fpresentation%2F1571098105.pdf
Publication
https://doi.org/10.1109/ICAIIC64266.2025.10920654

[E02] Enhancing LLMs for Power System Simulations: A Feedback-Driven Multi-Agent Framework

Authors / issuer
Mengshuo Jia, Zeyu Cui, Gabriela Hug
Publisher / venue
IEEE Transactions on Smart Grid, volume 16, issue 6, pages 5556–5572
Date
Published online 2025-07-18; issue November 2025; inspected full arXiv v3 dated 2025-05-19
Evidence type
Published original research; publisher metadata and full author preprint inspected
Full source URL
https://doi.org/10.1109/TSG.2025.3589114
Full paper PDF
https://arxiv.org/pdf/2411.16707v3

[E03] Grid-Agent: An LLM-Powered Multi-Agent System for Power Grid Control

Authors / issuer
Yan Zhang, Ahmad Mohammad Saber, Amr Youssef, Deepa Kundur
Publisher / venue
arXiv preprint
Date
2025-09-08 (v3); first submitted 2025-08-07
Evidence type
Original research; full v3 PDF
Full source URL
https://arxiv.org/pdf/2508.05702v3

[P02] PowerBench: A Benchmark for Agentic Retrieval and Reasoning in Power Systems

Authors / issuer
Xijing Wang, Yinsheng Yao, Jinru Ding, Yidong Jiang, Ziwen Xu, Yiwen Jiang, Jie Xu, Dawei Cheng
Publisher / venue
arXiv
Date
2026-09-28 (v1)
Type
Primary research preprint; synthetic retrieval benchmark
Location
Tables 2–4, PDF pp.5,7–8; Appendix B, D and F define tasks, metrics and limitations.
Full source URL
https://arxiv.org/abs/2609.34492v1
Full paper PDF
https://arxiv.org/pdf/2609.34492v1

[I04] Image published by NVIDIA with its EPRI Open Power AI Consortium announcement, March 20, 2025: Transmission towers and power lines with city lights in the background at dusk

Authors / issuer
NVIDIA
Publisher / venue
NVIDIA
Date
Image accessed 17 October 2026
Full source URL
https://blogs.nvidia.com/blog/open-power-ai-consortium/

[PA01] PowerAgent: A Road Map Toward Agentic Intelligence in Power Systems: Foundation Model, Model Context Protocol, and Workflow

Authors / issuer
Qian Zhang, Le Xie
Publisher / venue
IEEE Power and Energy Magazine, volume 23, issue 5, pages 93–101
Date
September–October 2025 issue; TechRxiv preprint DOI version 1, June 2025
Evidence type
Primary roadmap article; identity verified through Harvard author pages and indexed author preprint; full article download unavailable during this review
Full source URL
https://doi.org/10.1109/MPE.2025.3579718

[PA03] A Roadmap Towards Agentic Intelligence in Power Systems

Authors / issuer
Le Xie; slides credit Le Xie and Qian Zhang
Publisher / venue
ERCOT Technology and Security Committee guest presentation
Date
8 December 2025
Evidence type
Full original 19-page author presentation inspected; roadmap and workflow demonstrations
Full source URL
https://www.ercot.com/files/docs/2025/12/01/4-Guest-Presentation-A-Roadmap-toward-Agentic-Intelligence-in-Power-Systems.pdf

[P01] PowerAgentBench-Dyn: A Benchmark for Agentic AI in Power System Dynamic Studies

Authors / issuer
Qian Zhang, Andrea Pomarico, Costas Mylonas, Magda Foti, Alberto Berizzi, Le Xie
Publisher / venue
arXiv
Date
2026-06-18 (v1)
Type
Primary research preprint; prototype benchmark
Location
DMQ Tables I–II, PDF p.3; DSR Tables III–VI, pp.4–5.
Full source URL
https://arxiv.org/abs/2606.20401v1
Full paper PDF
https://arxiv.org/pdf/2606.20401v1

[PA06] PowerWF Grid Impact Evaluation — LangGraph workflow source

Authors / issuer
PowerAgent / PowerWF community contributors
Publisher / venue
Author-maintained source code, Grid Impact Evaluation/src/agent/graph.py
Date
Live source inspected 17 October 2026
Evidence type
Primary implementation: StateGraph, ToolNode, PowerWorld tool groups, conditional routing; inspected, not executed
Full source URL
https://github.com/Power-Agent/PowerWF/blob/main/Grid%20Impact%20Evaluation/src/agent/graph.py

[PA05] Power-Agent/PowerWF

Authors / issuer
PowerAgent community contributors
Publisher / venue
Author-maintained GitHub repository
Date
Live source repository reviewed 17 October 2026
Evidence type
Primary workflow source repository
Full source URL
https://github.com/Power-Agent/PowerWF

[P03] GridMind: An Agentic Workflow for Weather-Driven Power Grid Risk Assessment

Authors / issuer
D. Adrian Maldonado, Jonghwan Kwon, Sunash B. Sharma, Hongwei Jin, Stefano Fenu, Emil Constantinescu, David Sehloff, Parfait Gasana, Duane Rudolph Verner, Kibaek Kim
Publisher / venue
KDD 2026 Workshop on Scientific and Societal Agents / OpenReview
Date
August 2026 workshop; reviewed 17 October 2026
Type
Primary workshop paper; simulated workflow evaluation
Location
Architecture §3, PDF p.2; seven-query trace and ablation Table1, PDF p.4.
Full source URL
https://openreview.net/forum?id=Cs5TBzzNym
Full paper PDF
https://openreview.net/pdf/5ad1c2afd7298f9e6ce3bc25f18e7165555e1ffd.pdf

[VID01] Strategies for debugging agents with LangGraph Studio

Authors / issuer
LangChain
Publisher / venue
LangChain / YouTube
Date
26 March 2025
Evidence type
Official software-platform demonstration
Full source URL
https://www.youtube.com/watch?v=5vEC0Y4sV8g

[LG02] LangGraph Persistence

Authors / issuer
LangChain documentation maintainers
Publisher / venue
Official LangChain documentation
Date
Accessed 2026-10-17; live documentation
Evidence type
Official Python runtime documentation
Full source URL
https://docs.langchain.com/oss/python/langgraph/persistence

[LG05] Checkpointers

Authors / issuer
LangChain documentation maintainers
Publisher / venue
Docs by LangChain, Python LangGraph documentation
Date
Undated live documentation; accessed October 18, 2026
Type
Official framework persistence documentation and illustrations
Location
Opening Checkpoints illustration; Get state history / State illustration
Full source URL
https://docs.langchain.com/oss/python/langgraph/checkpointers

[PD02] DistGrid-AgentBench: A Distribution Grid Benchmark for Agentic AI

Authors / issuer
Emmanuel O. Badmus and contributors
Publisher / venue
Author-maintained GitHub repository
Date
Live repository reviewed 17 October 2026
Evidence type
Primary benchmark tasks, tool catalog and evaluator; README explicitly documents proprietary omissions
Full source URL
https://github.com/emmanuelbadmus/DistGrid-AgentBench

[R02] Agentic Artificial Intelligence for Power Systems: Strategies to Identify and Close Capability Gaps

Authors / issuer
Eve Tsybina, Samim Konjicija, Slaven Peles
Publisher / venue
arXiv
Date
30 July 2026; v1
Type
Research preprint; results apply to the tested prototype and suite
Location
Six task-complexity levels and four grid scales; only the two lowest levels solvable on some tested sizes
Full source URL
https://arxiv.org/abs/2607.28822v1
Full paper PDF
https://arxiv.org/pdf/2607.28822v1

[VID02] Gridscale X PSS E: agentic transmission planning — product video

Authors / issuer
Siemens
Publisher / venue
Siemens
Date
Undated video; accessed 26 October 2026
Evidence type
Official platform video; not evidence of a named utility deployment
Full source URL
https://video.sw.siemens.com/embed/682ca322-5d0b-47a0-8d18-77bf9d0fb016
Publication
https://www.siemens.com/en-gb/products/pss-software/gridscale-x-pss-e/agentic-transmission-planning/

[P04] GridMind: LLMs-Powered Agents for Power System Analysis and Operations

Authors / issuer
Hongwei Jin, Kibaek Kim, Jonghwan Kwon
Publisher / venue
arXiv; published as SC Workshops 2025, DOI 10.1145/3731599.3767409
Date
2025-09-02 (arXiv v1); SC Workshops 2025
Type
Primary research paper; early prototype
Location
§4, Fig3 and Table1, PDF pp.4–5.
Full source URL
https://arxiv.org/abs/2509.02494v1
Full paper PDF
https://arxiv.org/pdf/2509.02494v1

[P03A] GridAI — AI Foundation Models & Agents for the Power Grid

Authors / issuer
Argonne GridAI project team
Publisher / venue
Argonne GridAI project team
Date
Live project page reviewed 17 October 2026
Type
Author-maintained project and publication index
Full source URL
https://argonne-gridfm.github.io/

[P01C] Power-Agent/PowerAgentBench

Authors / issuer
PowerAgent Community
Publisher / venue
PowerAgent Community
Date
Live author repository reviewed 17 October 2026
Type
Author-maintained benchmark code, cases and specifications
Location
Dynamic level1 harness and README; licensed PSS/E and DMView prerequisites.
Full source URL
https://github.com/Power-Agent/PowerAgentBench

[E04] VeraGrid-Agent: Tool-Augmented LLMs for Distribution Optimal Power Flow at the Grid Edge

Authors / issuer
Shivanshu Tripathi, Hamed Mohsenian-Rad, Maziar Raissi
Publisher / venue
arXiv preprint
Date
2026-07-28 (v1)
Evidence type
Original research; full v1 PDF
Full source URL
https://arxiv.org/pdf/2607.25155v1

[E04C] Shivanshu-11/veragrid-llm-benchmark

Authors / issuer
Shivanshu Tripathi / VeraGrid-Agent authors
Publisher / venue
Author GitHub repository
Date
Reviewed 2026-10-17
Evidence type
Author implementation, benchmark and recorded-run repository
Full source URL
https://github.com/Shivanshu-11/veragrid-llm-benchmark

[R03] 7.2.4.1. Confidence intervals

Authors / issuer
National Institute of Standards and Technology, SEMATECH
Publisher / venue
NIST/SEMATECH e-Handbook of Statistical Methods
Date
Undated live handbook section; reviewed 17 October 2026
Type
Official statistical methods reference
Location
Wilson interval formula and exact binomial intervals. Tutorial intervals are calculated from R02 counts, not reported by the grid paper.
Full source URL
https://www.itl.nist.gov/div898/handbook/prc/section2/prc241.htm

[R04] EXACT BINOMIAL

Authors / issuer
National Institute of Standards and Technology
Publisher / venue
NIST Dataplot Reference Manual
Date
Last updated 16 November 2015; reviewed 17 October 2026
Type
Official statistical methods reference
Location
One-sided and two-sided exact binomial confidence limits. The tutorial derives the zero-failure bound from the binomial probability, with independent fixed-setting trials as an explicit assumption.
Full source URL
https://www.itl.nist.gov/div898/software/dataplot/refman2/auxillar/exacbino.htm

[T08] MCP Architecture overview

Authors / issuer
Model Context Protocol
Publisher / venue
Model Context Protocol
Date
2026-07-28 documentation revision
Evidence type
Official protocol documentation
Full source URL
https://modelcontextprotocol.io/docs/2026-07-28/learn/architecture

[T09] MCP Security Best Practices

Authors / issuer
Model Context Protocol
Publisher / venue
Model Context Protocol
Date
2026-07-28 documentation revision
Evidence type
Official protocol documentation
Full source URL
https://modelcontextprotocol.io/docs/2026-07-28/tutorials/security/security_best_practices

[U01] AES: Scaling energy safety audits with Claude on Google Cloud

Authors / issuer
Google Cloud
Publisher / venue
Google Cloud
Date
Undated; reviewed 17 October 2026
Type
vendor primary customer case study with named customer testimony
Full source URL
https://cloud.google.com/customers/aes

[U04] TC Energy increases office productivity with automation on Amazon Bedrock

Authors / issuer
Amazon Web Services
Publisher / venue
Amazon Web Services
Date
Undated; reviewed 17 October 2026
Type
vendor primary customer case study with named customer testimony
Full source URL
https://aws.amazon.com/solutions/case-studies/tcenergy-bedrock/

[O03] ISO now has expanded training and emergency operations facilities

Authors / issuer
Tricia Johnstone, Director, Operational Readiness
Publisher / venue
California ISO
Date
2025-02-20
Type
Official operator organization article
Full source URL
https://www.caiso.com/about/news/energy-matters-blog/iso-now-has-expanded-training-and-emergency-operations-facilities

[C15] CESER Releases New Testbed to Advance LLM and Agentic AI Evaluation for Critical Infrastructure

Authors / issuer
U.S. Department of Energy
Publisher / venue
U.S. Department of Energy
Date
2026-07-16
Type
government primary research-tool announcement
Full source URL
https://www.energy.gov/ceser/articles/ceser-releases-new-testbed-advance-llm-and-agentic-ai-evaluation-critical

[C11] EPRI Launches Effort to Reduce Risk in AI Power Sector Deployment

Authors / issuer
EPRI via GlobeNewswire
Publisher / venue
EPRI via GlobeNewswire
Date
2026-08-06
Type
issuer-authored primary press release
Full source URL
https://www.globenewswire.com/news-release/2026/08/06/3340219/0/en/epri-launches-effort-to-reduce-risk-in-ai-power-sector-deployment.html

[C12] SAFERai.power

Authors / issuer
EPRI
Publisher / venue
EPRI
Date
Undated; reviewed 17 October 2026
Type
primary initiative and roadmap
Full source URL
https://msites.epri.com/saferai

[I05] Oak Ridge National Laboratory, U.S. Department of Energy: Researchers at ORNL GOAL facing a large display of U.S. grid and geographic data

Authors / issuer
Oak Ridge National Laboratory, U.S. Department of Energy
Publisher / venue
Oak Ridge National Laboratory, U.S. Department of Energy
Date
Image accessed 17 October 2026
Full source URL
https://www.ornl.gov/grid-operations-and-analytics-laboratory-goal/

[L02] Existing three-use-case GridGuard workshop: instructor route and synthetic-feeder context

Authors / issuer
Local workshop source snapshot
Publisher / venue
Local workshop source snapshot
Date
Snapshot 17 October 2026
Location
Voltage support, service restoration and cyber response are separate worked modules.
Full source URL
research/Follow_on_workshop_source.txt

[P02C] open-compass/PowerBench

Authors / issuer
OpenCompass / PowerBench authors
Publisher / venue
OpenCompass / PowerBench authors
Date
Live author-linked repository reviewed 17 October 2026
Type
Linked project repository; only README and LICENSE visible at review
Full source URL
https://github.com/open-compass/PowerBench

[P03M] KDD 2026 SciSoc LLM Workshop — Accepted Papers

Authors / issuer
KDD26SciSocLLM workshop organizers
Publisher / venue
KDD26SciSocLLM workshop organizers
Date
2026 workshop listing
Type
Official workshop metadata identifying GridMind authors and forum
Full source URL
https://github.com/KDD26SciSocLLM/KDD26SciSocLLM.github.io/blob/master/accepted-papers.md

[PA04] Power-Agent/PowerMCP

Authors / issuer
PowerAgent community contributors
Publisher / venue
Author-maintained GitHub repository
Date
Live source repository reviewed 17 October 2026
Evidence type
Primary implementation and integration documentation
Full source URL
https://github.com/Power-Agent/PowerMCP

Visual credit: real_framework_nvidia_interaction_patterns.svg

Caption
Original NVIDIA interaction patterns. The same generative model can be part of a chat or an agentic system.
Credit
Source: NVIDIA Technical Blog, Figure 1.
Used on
2. What is agentic AI?
Transformation
Only blank outer margins trimmed by display viewport; original image bytes are embedded unchanged and also retained as a separate original file.
Evidence limit
The three examples show interaction patterns, not a universal taxonomy or a claim that tool use alone proves autonomy.
License or permission status
Publisher attribution supplied. No separate permission or unrestricted reuse license claimed.
Original source / page
https://developer.nvidia.com/blog/building-for-the-rising-complexity-of-agentic-systems-with-extreme-co-design/
Original figure / image URL
https://developer-blogs.nvidia.com/wp-content/uploads/2026/05/image7.webp

Visual credit: real_concept_classical_agent.png

Caption
Russell and Norvig’s original agent diagram, Figure 2.1. A software agent’s sensors and actuators can be software interfaces.
Credit
Stuart J. Russell and Peter Norvig, Artificial Intelligence: A Modern Approach, second edition, Chapter 2, Figure 2.1, printed p.33; author-hosted chapter draft.
Used on
4. What is an AI agent?
Asset type
Original published textbook figure, native PDF crop
Transformation
Native PDF page clipped around Figure 2.1; no figure labels or connections redrawn.
Evidence limit
A general agent model, not a utility deployment or an LLM-specific architecture.
Original source / page
https://aima.cs.berkeley.edu/4th-ed/pdfs/newchap02.pdf
Original figure / image URL
https://aima.cs.berkeley.edu/4th-ed/pdfs/newchap02.pdf

Visual credit: real_framework_nvidia_agent_components.svg

Caption
Original NVIDIA agent design: the core connects a request with tools, planning and memory. Other designs can combine these roles.
Credit
Source: NVIDIA Technical Blog, Figure 1.
Used on
5. An agent chooses a step. A tool runs it.
Transformation
Only blank outer margins trimmed by display viewport; original image bytes are embedded unchanged and also retained as a separate original file.
Evidence limit
This is a published conceptual design; separate memory or planning modules are not required in every AI agent.
License or permission status
Publisher attribution supplied. No separate permission or unrestricted reuse license claimed.
Original source / page
https://developer.nvidia.com/blog/introduction-to-llm-agents/
Original figure / image URL
https://developer-blogs.nvidia.com/wp-content/uploads/2023/11/agent-components.png

Visual credit: concept_Qd6anWv0mv0.jpg

Caption
Official Google Cloud Tech concept video. The grid mapping is an original teaching analogy.
Credit
Google Cloud Tech. Aja Hammerly and Jason Davenport. Agentic AI: Workflows vs. agents. 19 March 2025. Official video thumbnail.
Used on
6. Watch: how an agent chooses its next step
Asset type
Official video thumbnail
Transformation
None
Evidence limit
Concept explainer with general software examples. The operator workflow on this slide is an original teaching analogy.
License or permission status
Publisher thumbnail linked to the official video; copyright remains with the publisher.
Original source / page
https://www.youtube.com/watch?v=Qd6anWv0mv0
Original figure / image URL
https://i.ytimg.com/vi/Qd6anWv0mv0/maxresdefault.jpg

Visual credit: real_framework_langgraph_review_detail.svg

Caption
Original LangGraph Studio review-form excerpt. This is a generic demo, not a utility interface.
Credit
Source: LangChain documentation maintainers, official LangGraph documentation. Accessed October 18, 2026.
Used on
7. An agent needs a clear goal and limits
Transformation
Viewport-only excerpt of the original approval form. Original image bytes embedded unchanged. Surrounding request and unused page whitespace excluded; no interface values edited.
Evidence limit
This is a generic software demonstration, not an operator dispatch interface or a utility deployment.
License or permission status
Publisher attribution supplied. No separate permission or unrestricted reuse license claimed.
Original source / page
https://docs.langchain.com/oss/python/langgraph/interrupts
Original figure / image URL
https://mintcdn.com/langchain-5e9cc07a/u62GP5wl5NeQqwjE/images/langgraph-interrupt-response-schema.png

Visual credit: source_anthropic_augmented_llm_wide.svg

Caption
Original Anthropic building-block figure. It is a general pattern, not a diagram of PowerAgent or PowerDAG.
Credit
Source: Anthropic, Building effective agents (2024), The augmented LLM. Original published figure, original diagram pixels unchanged; white margins trimmed for display.
Used on
8. A model can use records, tools and memory
Transformation
White-margin-only display trim: SVG viewport clips an embedded byte-identical original PNG; PDF places original image samples on a cropped page. No diagram pixels or labels edited. The one-pixel outer frame is excluded.
Evidence limit
Published conceptual pattern, not an implementation diagram of PowerAgent, PowerDAG or a utility deployment.
License or permission status
No explicit image license verified; attribution is supplied, not a claim of an open license.
Original source / page
https://www.anthropic.com/engineering/building-effective-agents
Original figure / image URL
https://www-cdn.anthropic.com/images/4zrzovbb/website/d3083d3f40bb2b6f477901cc9a240738d3dd1371-2401x1000.png

Visual credit: real_news_pjm_hyperq_header.png

Caption
Original Tapestry article and published animation still. Supplier report, June 2026.
Credit
Tapestry, HyperQ Field Notes, June 11, 2026
Used on
9. PJM uses HyperQ to review connection applications
Asset type
Genuine browser screenshot of original official source page
Transformation
Viewport crop of the live original webpage only. No text reconstruction, added labels, retouching, content changes or synthetic card.
Evidence limit
Supplier report of deployment. PJM separately confirmed earlier testing. The hero is a still captured from the publisher's animation, not an independently verified live PJM application record.
License or permission status
Original publisher content. No explicit open-reuse license independently verified. Attribution records the source and does not assert permission.
Original source / page
https://www.tapestryenergy.com/en/projects/field-notes-hyperq-deployment-in-pjm-s-cycle-1

Visual credit: power_news_hyperq_dashboard.png

Caption
Original Tapestry performance graphic, June 2026. The timing covers only computer processing.
Credit
Tapestry, HyperQ field notes, June 11, 2026
Used on
10. HyperQ: what the reported numbers measure
Asset type
Original publisher site-control performance infographic; not an application-report screenshot
Transformation
None: original static PNG bytes, unmodified; no cropping, resizing, retouching, or GIF-frame conversion.
Evidence limit
Supplier-reported deployment metrics. Processing time excludes document transmission and human validation. 9,312 macro-assessments are not 9,312 applications.
Original source / page
https://www.tapestryenergy.com/en/projects/field-notes-hyperq-deployment-in-pjm-s-cycle-1
Original figure / image URL
https://cdn.sanity.io/images/bhywzdze/production/ee934ed00f306c6f3c2d0b5a92ab78dcb6278584-1257x651.png

Visual credit: caiso_control_room.jpg

Caption
Real California ISO control room. OATI supplies the batch result and later production-use claim.
Credit
California ISO control room, image published by OATI
Used on
11. CAISO/OATI: agents help review planned outages
Asset type
Photograph
Transformation
None
Evidence limit
Control-room context photograph. The photo does not establish that any pictured display shows an AI agent or that an agent controls grid equipment.
License or permission status
Official OATI case-study page identifies this as the California ISO control room. No image-specific photographer or open reuse license stated. Permission not independently obtained.
Original source / page
https://www.oati.com/ai-genie/
Original figure / image URL
https://www.oati.com/wp-content/uploads/FolsomControlRoom5x7-1-1-2000x1435.jpg

Visual credit: duke_operator.jpg

Caption
Duke operator, 2022, shown for context. The reported agents support planning, rather than this pictured operating desk.
Credit
Duke Energy, illumination, August 2, 2022
Used on
12. Duke and AWS: agents help prepare planning studies
Asset type
Photograph
Transformation
None
Evidence limit
Historical operator photograph used for operating context. The 2022 story does not document agentic AI.
License or permission status
Duke Energy Corporation, all rights reserved according to source-page footer. Permission not independently obtained.
Original source / page
https://illumination.duke-energy.com/articles/meet-the-team-that-monitors-the-grid-and-keeps-the-lights-on
Original figure / image URL
https://iprsoftwaremedia.com/264/files/20227/2022-0802-ECC-Greene-5.jpg

Visual credit: real_news_siemens_gridscale_header.png

Caption
Original Siemens press-release header, 2026. This is an announced product capability.
Credit
Siemens AG, press release, May 5, 2026
Used on
13. Siemens announces agents for PSS E planning studies
Asset type
Faithful rendering of original official PDF page
Transformation
Original PDF page rendered by Poppler at aspect-preserving 1800px maximum dimension. No content edited. Source-preserving rectangular pixel crop [0, 0, 1273, 485]. Original text/image pixels unchanged; omitted surrounding page body, form or cookie-consent areas.
Evidence limit
Product announcement. The separately named Alliander platform migration does not prove production use of the new agentic PSS E capability.
License or permission status
Original publisher content. No explicit open-reuse license independently verified. Attribution records the source and does not assert permission.
Original source / page
https://press.siemens.com/global/en/pressrelease/siemens-gridscale-x-redefines-system-operations-and-agentic-transmission-planning
Original figure / image URL
https://assets.new.siemens.com/siemens/assets/api/uuid:fb613ac9-42b5-4e56-85be-0881ab9e95b8/HQSIPR202604177385EN.pdf

Visual credit: real_news_nypa_webinar_detail.svg

Caption
Original Ataccama webinar artwork. The NYPA speaker appears at lower right.
Credit
Ataccama, original webinar report with NYPA Data Quality Lead Saptaparna Gupta, October 14, 2026
Used on
15. NYPA: better data for agent use
Asset type
Genuine browser screenshot of original official source page
Transformation
Viewport-only excerpt of the original webinar artwork. Page heading/date remain in the retained original screenshot; no artwork text or portraits changed.
Evidence limit
NYPA discusses meter-data quality checks. The separate product/reorder demo does not establish a production NYPA grid-control agent.
License or permission status
Original publisher content. No explicit open-reuse license independently verified. Attribution records the source and does not assert permission.
Original source / page
https://www.ataccama.com/webinar/on-demand-trust-driven-medallion-architecture

Visual credit: real_news_eversource_job_header.png

Caption
Original Eversource vacancy, 5 October 2026.
Credit
Eversource Energy, official engineering vacancy, October 5, 2026
Used on
16. Eversource: adding agentic skills to grid studies
Asset type
Genuine browser screenshot of original official source page
Transformation
Viewport crop of the live original webpage only. No text reconstruction, added labels, retouching, content changes or synthetic card. Source-preserving rectangular pixel crop [0, 0, 1440, 875]. Original text/image pixels unchanged; omitted surrounding page body, form or cookie-consent areas.
Evidence limit
Engineering skill requirement and integration intent. The posting later explicitly mentions agentic MATLAB workflows, but a vacancy does not prove deployment.
License or permission status
Original publisher content. No explicit open-reuse license independently verified. Attribution records the source and does not assert permission.
Original source / page
https://jobs.eversource.com/job/east-berlin/distribution-system-modeling-and-automation-engineer-hybrid-schedule/7774/101576695504

Visual credit: real_news_eversource_agentic_line.svg

Caption
Original requirement excerpt. Hiring and integration intent; deployment is not established.
Credit
Eversource Energy, official engineering vacancy, October 5, 2026
Used on
16. Eversource: adding agentic skills to grid studies
Asset type
Genuine browser element screenshot of original published job requirements
Transformation
Viewport selects the single complete agentic-workflow requirement from the original source-element screenshot. Original image bytes retained unchanged.
Evidence limit
Advertised engineering skills. This is hiring/integration intent, not proof of a production agent platform.
License or permission status
Original publisher content. No explicit open-reuse license independently verified. Attribution records the source and does not assert permission.
Original source / page
https://jobs.eversource.com/job/east-berlin/distribution-system-modeling-and-automation-engineer-hybrid-schedule/7774/101576695504

Visual credit: ge_gridos_control_room.jpg

Caption
Official GridOS promotional composite. The platform example does not establish a measured agent-specific result.
Credit
GE Vernova
Used on
17. GE Vernova: a platform that can connect grid tools
Asset type
Official product promotional composite
Transformation
Format conversion from WebP to JPEG only
Evidence limit
Promotional product imagery, not evidence that an autonomous agent operates a named U.S. utility control room.
License or permission status
Official GE Vernova product-page image. No explicit reuse license found. Permission not independently obtained.
Original source / page
https://www.gevernova.com/software/products/gridos
Original figure / image URL
https://www.gevernova.com/software/sites/default/files/2026-06/2024-grid-gridos-hero-image-1600x536.webp

Visual credit: duke_control_room.jpg

Caption
Duke Energy control room, 2022. The PNNL survey is a separate study.
Credit
Duke Energy, illumination, August 2, 2022
Used on
18. Operators often work against a deadline
Asset type
Photograph
Transformation
None
Evidence limit
Historical operator photograph used for operating context. The 2022 story does not document agentic AI.
License or permission status
Duke Energy Corporation, all rights reserved according to source-page footer. Permission not independently obtained.
Original source / page
https://illumination.duke-energy.com/articles/meet-the-team-that-monitors-the-grid-and-keeps-the-lights-on
Original figure / image URL
https://iprsoftwaremedia.com/264/files/20227/62e806b2b3aed342060cf206_2022-0802-ECC-Greene-hero/2022-0802-ECC-Greene-hero_hero.jpg?v=b4580dcf-c141-4a6a-b29f-a5038a74cb1a

Visual credit: caiso_watch_duty.png

Caption
Real CAISO Watch Duty display. It combines fire and transmission-line information for operator awareness.
Credit
California ISO. Application: Watch Duty.
Used on
19. Three checks for a useful recommendation
Asset type
Actual operational situational-awareness tool screenshot
Transformation
Format conversion from WebP to PNG only
Evidence limit
The source describes operators overlaying transmission lines on fire maps and reviewing webcam images to assess fire proximity. This screenshot is not agentic AI and does not establish autonomous control. Native source resolution is modest, so avoid large enlargement.
License or permission status
Official California ISO control-center article screenshot. Source footer states all rights reserved. No image-specific reuse license found. Permission not independently obtained.
Original source / page
https://www.caiso.com/about/news/energy-matters-blog/control-center-modernization-is-enhancing-reliability
Original figure / image URL
https://www.caiso.com/content/images/blog/blog-graphics/controlcentermodernization-watchduty.webp

Visual credit: ge_gridos_visual_intelligence.png

Caption
GE Vernova’s published GridOS Visual Intelligence screen. This product image supplies task context, not an agent result.
Credit
GE Vernova
Used on
20. A clear task makes success easier to test
Asset type
Official product screenshot and video thumbnail
Transformation
None
Evidence limit
Illustrates vendor software, not a named U.S. agentic AI deployment.
License or permission status
Official GE Vernova product-page video thumbnail. No explicit reuse license found. Permission not independently obtained.
Original source / page
https://www.gevernova.com/software/products/gridos/visual-intelligence
Original figure / image URL
https://www.gevernova.com/software/sites/default/files/2026-10/frame-1171277779.png

Visual credit: source_anthropic_autonomous_agent_wide.svg

Caption
Original Anthropic agent-loop figure. In a grid study, feedback can be solver results or a failed check.
Credit
Source: Anthropic, Building effective agents (2024), Autonomous agent. Original published figure, original diagram pixels unchanged; white margins trimmed for display.
Used on
21. An LLM agent acts, reads the result and decides again
Transformation
White-margin-only display trim: SVG viewport clips an embedded byte-identical original PNG; PDF places original image samples on a cropped page. No diagram pixels or labels edited. The one-pixel outer frame is excluded.
Evidence limit
Published conceptual pattern, not an implementation diagram of PowerAgent, PowerDAG or a utility deployment.
License or permission status
No explicit image license verified; attribution is supplied, not a claim of an open license.
Original source / page
https://www.anthropic.com/engineering/building-effective-agents
Original figure / image URL
https://www-cdn.anthropic.com/images/4zrzovbb/website/58d9f10c985c4eb5d53798dea315f7bb5ab6249e-2401x1000.png

Visual credit: real_framework_anthropic_prompt_chain.svg

Caption
Original Anthropic gated workflow. In our study example, a stale input should fail the check and stop the sequence.
Credit
Source: Anthropic, Building effective agents (December 19, 2024).
Used on
22. When data are stale, the next step is to stop
Transformation
Only blank outer margins trimmed by display viewport; original image bytes are embedded unchanged and also retained as a separate original file.
Evidence limit
A generic content-workflow pattern, not a physical-feasibility or utility-safety guarantee.
License or permission status
Publisher attribution supplied. No separate permission or unrestricted reuse license claimed.
Original source / page
https://www.anthropic.com/engineering/building-effective-agents
Original figure / image URL
https://www-cdn.anthropic.com/images/4zrzovbb/website/7418719e3dab222dccb379b8879e1dc08ad34c78-2401x1000.png

Visual credit: real_framework_anthropic_routing.svg

Caption
Original Anthropic routing pattern. Choosing an engineering study also requires selecting the right problem and tool.
Credit
Source: Anthropic, Building effective agents (December 19, 2024).
Used on
25. A solver result depends on the problem it solves
Transformation
Only blank outer margins trimmed by display viewport; original image bytes are embedded unchanged and also retained as a separate original file.
Evidence limit
The picture illustrates routing; it does not establish an implemented power-system workflow.
License or permission status
Publisher attribution supplied. No separate permission or unrestricted reuse license claimed.
Original source / page
https://www.anthropic.com/engineering/building-effective-agents
Original figure / image URL
https://www-cdn.anthropic.com/images/4zrzovbb/website/5c0c0e9fe4def0b584c04d37849941da55e5e71c-2401x1000.png

Visual credit: epri_open_power_ai_grid.png

Caption
Image accompanying the EPRI Open Power AI Consortium announcement, published by NVIDIA, 20 March 2025.
Credit
Image published by NVIDIA with its EPRI Open Power AI Consortium announcement, March 20, 2025
Used on
28. Power-system papers test different tasks
Asset type
Photograph
Transformation
None
Evidence limit
Contextual grid photograph. The pictured location and utility are not identified.
License or permission status
Official NVIDIA blog image. No photographer or explicit reuse license stated on the source page. Permission not independently obtained.
Original source / page
https://blogs.nvidia.com/blog/open-power-ai-consortium/
Original figure / image URL
https://blogs.nvidia.com/wp-content/uploads/2025/03/cudss-featured-1280x680-blog.png

Visual credit: real_paper_poweragent_three_layers.png

Caption
Original Harvard author roadmap: workflows, tool connections and foundation models. Three-layer excerpt.
Credit
Le Xie; slides credit Le Xie and Qian Zhang; A Roadmap Towards Agentic Intelligence in Power Systems; Author presentation slide 11
Used on
29. Harvard PowerAgent: models, tools and study workflows
Asset type
Original published figure or result panel, native PDF crop
Transformation
Native PDF crop selects the three layer panels from the author roadmap. No text or connectors redrawn; the full source figure is retained.
Evidence limit
Roadmap from the December2025 author presentation; not a measured deployment or a frozen specification of every current PowerAgent repository.
Original source / page
https://www.ercot.com/files/docs/2025/12/01/4-Guest-Presentation-A-Roadmap-toward-Agentic-Intelligence-in-Power-Systems.pdf
Original figure / image URL
https://www.ercot.com/files/docs/2025/12/01/4-Guest-Presentation-A-Roadmap-toward-Agentic-Intelligence-in-Power-Systems.pdf

Visual credit: real_paper_poweragent_impact_workflow.png

Caption
Original Harvard concept for grid-impact studies. The separate PowerWF example uses LangGraph and PowerWorld; this is not its exact graph.
Credit
Le Xie; slides credit Le Xie and Qian Zhang; A Roadmap Towards Agentic Intelligence in Power Systems; Author presentation slide 14, right-hand grid-impact process
Used on
30. PowerAgent example: what changes when load is added?
Asset type
Original published figure or result panel, native PDF crop
Transformation
Only the stated page rectangle is retained; original labels, plotted content and colors are unchanged. Non-figure page-link annotations are removed. Native PDF content is preserved, with a 240-dpi PNG for web display.
Evidence limit
A conceptual author demonstration. It is not an exact diagram of the later PowerWF graph.py or a utility performance trial.
Original source / page
https://www.ercot.com/files/docs/2025/12/01/4-Guest-Presentation-A-Roadmap-toward-Agentic-Intelligence-in-Power-Systems.pdf
Original figure / image URL
https://www.ercot.com/files/docs/2025/12/01/4-Guest-Presentation-A-Roadmap-toward-Agentic-Intelligence-in-Power-Systems.pdf

Visual credit: power_dragent_waveform.png

Caption
Original DrAgent Figure 2, PDF p.4. The third current channel shows a fault in a sample disturbance record.
Credit
Saha, V and Naidu / Hitachi Energy, DrAgent, ICAIIC 2025, Figure 2
Used on
31. DrAgent: reading a recorded grid fault
Asset type
Original scientific figure, native PDF crop
Transformation
Clipped to the original figure boundary; caption excluded and credited separately; no plotted data changed.
Evidence limit
A sample disturbance record from a research paper; not a live utility operating display.
Original source / page
https://www.manuscriptlink.com/society/kics/media?key=kics%2Fconference%2Ficaiic2025%2Fpresentation%2F1571098105.pdf
Original figure / image URL
https://www.manuscriptlink.com/society/kics/media?key=kics%2Fconference%2Ficaiic2025%2Fpresentation%2F1571098105.pdf

Visual credit: real_paper_dragent_architecture.png

Caption
Original DrAgent architecture: planning, tool execution and answer generation.
Credit
Barun Kumar Saha, Aarthi V, O. D. Naidu; DrAgent: An Agentic Approach to Fault Analysis in Power Grids Using Large Language Models; Figure 1, PDF page 2
Used on
32. DrAgent selects tools to answer a fault question
Asset type
Original published figure or result panel, native PDF crop
Transformation
Only the stated page rectangle is retained; original labels, plotted content and colors are unchanged. Non-figure page-link annotations are removed. Native PDF content is preserved, with a 240-dpi PNG for web display.
Evidence limit
Paper architecture based on ReWOO; the picture does not establish utility deployment or autonomous command authority.
Original source / page
https://www.manuscriptlink.com/society/kics/media?key=kics%2Fconference%2Ficaiic2025%2Fpresentation%2F1571098105.pdf
Original figure / image URL
https://www.manuscriptlink.com/society/kics/media?key=kics%2Fconference%2Ficaiic2025%2Fpresentation%2F1571098105.pdf

Visual credit: power_dragent_counts.png

Caption
Counts reconstructed from Table IV and the Appendix. They are neither new experiments nor the paper’s mean-score measure.
Credit
Original educational plot derived from published DrAgent score cells; no experiments rerun.
Used on
33. DrAgent: more correct answers, with errors still present
Asset type
Original quantitative teaching figure
Provenance note
Correct answers (score >=2), reconstructed from Table IV and Appendix; n=18 DR cases per question. Q6 includes correctly labeled duration under footnote 8. Counts are author-data-derived, not an independent rerun.
Evidence limit
18 disturbance-record cases, not 18 repeated stochastic trials per case. Scores were manually assessed by a domain expert. Baseline and agent use different prompt/context workflows. The comparison does not isolate one component causally. Do not describe the Q6 result as 13/18 correct clearance timestamps. Q6 printed mean is inconsistent; audited cell counts are used instead.
Original source / page
https://www.manuscriptlink.com/society/kics/media?key=kics%2Fconference%2Ficaiic2025%2Fpresentation%2F1571098105.pdf
Original figure / image URL
images/power_dragent_counts.png

Visual credit: real_framework_langgraph_static_interrupt.svg

Caption
Real LangGraph debugger screenshot. A paused run can be inspected before the next software step.
Credit
Source: LangChain documentation maintainers, official LangGraph documentation. Accessed October 18, 2026.
Used on
36. LangGraph can pause to inspect a run
Transformation
Only blank outer margins trimmed by display viewport; original image bytes are embedded unchanged and also retained as a separate original file.
Evidence limit
A static debug breakpoint differs from a dynamic interrupt requesting human input.
License or permission status
Publisher attribution supplied. No separate permission or unrestricted reuse license claimed.
Original source / page
https://docs.langchain.com/oss/python/langgraph/interrupts
Original figure / image URL
https://mintcdn.com/langchain-5e9cc07a/dL5Sn6Cmy9pwtY0V/oss/images/static-interrupt.png

Visual credit: platform_5vEC0Y4sV8g.jpg

Caption
LangChain / 26 March 2025
Credit
LangChain. Strategies for debugging agents with LangGraph Studio. 26 March 2025. Official video thumbnail.
Used on
37. LangGraph demo: pause, inspect and continue
Asset type
Official video thumbnail
Transformation
None
Evidence limit
General software debugging demo. The selected classroom excerpt is 3:34–5:13; it does not demonstrate power-system feasibility or utility deployment.
License or permission status
Publisher thumbnail linked to the official video; copyright remains with the publisher.
Original source / page
https://www.youtube.com/watch?v=5vEC0Y4sV8g
Original figure / image URL
https://i.ytimg.com/vi/5vEC0Y4sV8g/hqdefault.jpg

Visual credit: real_framework_langgraph_state.svg

Caption
Original LangGraph state diagram. Software checkpoints still need grid-case, time and evidence identifiers.
Credit
Source: LangChain documentation maintainers, official LangGraph documentation. Accessed October 18, 2026.
Used on
38. Saved state must identify the study and its evidence
Transformation
Selected original figure panel excerpt; no labels, arrows or pixels edited. Entire original image retained separately.
Evidence limit
The extra grid identifiers are tutorial requirements; they are not fields proven by this generic illustration.
License or permission status
Publisher attribution supplied. No separate permission or unrestricted reuse license claimed.
Original source / page
https://docs.langchain.com/oss/python/langgraph/checkpointers
Original figure / image URL
https://mintcdn.com/langchain-5e9cc07a/-_xGPoyjhyiDWTPJ/oss/images/get_state.jpg

Visual credit: real_framework_langgraph_checkpoint.svg

Caption
Original LangGraph checkpoint diagram. Saving progress does not prove that the grid data still match.
Credit
Source: LangChain documentation maintainers, official LangGraph documentation. Accessed October 18, 2026.
Used on
39. A checkpoint lets a workflow pause and resume
Transformation
Selected original figure panel excerpt; no labels, arrows or pixels edited. Entire original image retained separately.
Evidence limit
Saved software state does not prove that an external grid snapshot remains current.
License or permission status
Publisher attribution supplied. No separate permission or unrestricted reuse license claimed.
Original source / page
https://docs.langchain.com/oss/python/langgraph/checkpointers
Original figure / image URL
https://mintcdn.com/langchain-5e9cc07a/-_xGPoyjhyiDWTPJ/oss/images/checkpoints.jpg

Visual credit: real_framework_anthropic_parallel.svg

Caption
Original Anthropic parallel workflow. A grid study must also account for missing or failed outage cases.
Credit
Source: Anthropic, Building effective agents (December 19, 2024).
Used on
40. Parallel outage studies must account for every case
Transformation
Only blank outer margins trimmed by display viewport; original image bytes are embedded unchanged and also retained as a separate original file.
Evidence limit
The aggregator picture does not imply that missing results are automatically detected.
License or permission status
Publisher attribution supplied. No separate permission or unrestricted reuse license claimed.
Original source / page
https://www.anthropic.com/engineering/building-effective-agents
Original figure / image URL
https://www-cdn.anthropic.com/images/4zrzovbb/website/406bb032ca007fd1624f261af717d70e6ca86286-2401x1000.png

Visual credit: real_paper_powerdag_architecture.png

Caption
Original PowerDAG v4 architecture, including the supervisor before environment execution.
Credit
Emmanuel O. Badmus, Amritanshu Pandey; PowerDAG: Supervisory Agentic AI System for Automating Distribution Grid Analysis; Figure 2, v4 PDF page 3
Used on
41. PowerDAG checks the steps needed before a tool runs
Asset type
Original published figure or result panel, native PDF crop
Transformation
Only the stated page rectangle is retained; original labels, plotted content and colors are unchanged. Non-figure page-link annotations are removed. Native PDF content is preserved, with a 240-dpi PNG for web display.
Evidence limit
This is the v4 research architecture. Formal supervision checks encoded prior Write-tool names; it does not establish all freshness, parameter or physical-safety checks.
Original source / page
https://arxiv.org/pdf/2603.17418v4
Original figure / image URL
https://arxiv.org/pdf/2603.17418v4

Visual credit: powerdag_revision_result.png

Caption
PowerDAG v4, Table II. ReAct: a reason-and-act loop. JIT: prerequisite checks just before a call.
Credit
Original vector plot redrawn from author-reported PowerDAG v4 numerical results
Used on
43. PowerDAG results: the same queries, four setups
Asset type
Original quantitative teaching figure
Provenance note
PowerDAG v4 same-model comparison on150 held-out distribution-grid queries: GPT-5.5 ReAct77.33%, adaptive-retrieval-only92.67%, JIT-supervision-only96.00%, fullPowerDAG98.00%. Each query is run once per model/configuration; no experiments rerun.
Evidence limit
Pass@1 checks expert tool coverage, compatible arguments and dependency ordering. Four real-world feeder models and ten task families; within-family holdout, not field reliability. No run-to-run interval is inferred. Public network-dependent reruns require proprietary data/tools; only3of10families are fully runnable publicly.
Original source / page
https://arxiv.org/pdf/2603.17418v4
Original figure / image URL
images/powerdag_revision_result.png

Visual credit: real_paper_daline_feedback.png

Caption
Original environmental acting module: simulation errors feed retrieval and code repair.
Credit
Mengshuo Jia, Zeyu Cui, Gabriela Hug; Enhancing LLMs for Power System Simulations: A Feedback-Driven Multi-Agent Framework; Figure 5, author-preprint PDF page 7
Used on
45. Simulation agents find functions, run code and repair errors
Asset type
Original published figure or result panel, native PDF crop
Transformation
Only the stated page rectangle is retained; original labels, plotted content and colors are unchanged. Non-figure page-link annotations are removed. Native PDF content is preserved, with a 240-dpi PNG for web display.
Evidence limit
Published workflow module, not evidence that error-free execution satisfies the requested engineering settings.
Original source / page
https://doi.org/10.1109/TSG.2025.3589114
Original figure / image URL
https://arxiv.org/pdf/2411.16707v3

Visual credit: power_study_feedback.png

Caption
DALINE comparison reported in the author preprint, Figure 8 / section V. Each task allows up to three attempts.
Credit
Original plot redrawn from cited primary-paper numerical results; no experiments rerun
Used on
46. Simulation-agent results: points, not a pass rate
Asset type
Original quantitative teaching figure
Provenance note
DALINE graded attempt-averaged scores: GPT4o-Sole 0.00, GPT4o-SR 31.37, CGPT4o-R 33.82 and GPT4o-Full 93.13. The evaluation uses 34 tasks with a maximum of three attempts each, awarding 100/50/0 points and carrying the last executed score into unused slots. This is not a final-completion fraction.
Evidence limit
Reported values preserved. API schemes use gpt-4o-2024-05-13; CGPT4o-R is the ChatGPT4o built-in-RAG product scheme. Complete-system comparison, not isolated causal effects. No binomial interval or implied completed-task count.
Original source / page
https://arxiv.org/pdf/2411.16707v3
Original figure / image URL
images/power_study_feedback.png

Visual credit: real_framework_anthropic_evaluator.svg

Caption
Original Anthropic feedback pattern. A retry must actually use the repaired code and changed inputs.
Credit
Source: Anthropic, Building effective agents (December 19, 2024).
Used on
47. Trying again can repeat the same software error
Transformation
Only blank outer margins trimmed by display viewport; original image bytes are embedded unchanged and also retained as a separate original file.
Evidence limit
An LLM evaluator does not prove grid feasibility, permission or fresh input data.
License or permission status
Publisher attribution supplied. No separate permission or unrestricted reuse license claimed.
Original source / page
https://www.anthropic.com/engineering/building-effective-agents
Original figure / image URL
https://www-cdn.anthropic.com/images/4zrzovbb/website/14f51e6406ccb29e695da48b17017e899a6119c7-2401x1000.png

Visual credit: platform_682ca322-5d0b-47a0-8d18-77bf9d0fb016.png

Caption
Siemens / Undated video; accessed 26 October 2026
Credit
Siemens. Gridscale X PSS E: agentic transmission planning — product video. Undated video; accessed 26 October 2026. Official video thumbnail.
Used on
48. Siemens PSS E: agentic transmission planning
Asset type
Official video thumbnail
Transformation
None
Evidence limit
General software-platform demonstration; not power-system validation or utility deployment evidence.
License or permission status
Publisher thumbnail linked to the official video; copyright remains with the publisher.
Original source / page
https://video.sw.siemens.com/embed/682ca322-5d0b-47a0-8d18-77bf9d0fb016
Original figure / image URL
https://video.sw.cdn.siemens.com/mgc/thumbnails/682ca322-5d0b-47a0-8d18-77bf9d0fb016-en-US-thumbnail.png

Visual credit: real_paper_gridmind2025_architecture.png

Caption
Original GridMind2025 diagram of specialized agents and numerical analysis tools.
Credit
Hongwei Jin, Kibaek Kim, Jonghwan Kwon; GridMind: LLMs-Powered Agents for Power System Analysis and Operations; Figure 2, PDF page 2
Used on
49. GridMind 2025 connects typed agents to solvers
Asset type
Original published figure or result panel, native PDF crop
Transformation
Only the stated page rectangle is retained; original labels, plotted content and colors are unchanged. Non-figure page-link annotations are removed. Native PDF content is preserved, with a 240-dpi PNG for web display.
Evidence limit
This is the2025 PydanticAI/pandapower study, not the separate2026 LangGraph weather-risk system.
Original source / page
https://arxiv.org/abs/2509.02494v1
Original figure / image URL
https://arxiv.org/pdf/2509.02494v1

Visual credit: real_paper_gridmind2025_response.png

Caption
A separate published response claims no thermal overloads and flags a balance error. It is not independent ground truth.
Credit
Hongwei Jin, Kibaek Kim, Jonghwan Kwon; GridMind: LLMs-Powered Agents for Power System Analysis and Operations; Figure 8, PDF page 10, baseline and coverage portion of agent response
Used on
50. GridMind: reported answers need independent checks
Asset type
Original published figure or result panel, native PDF crop
Transformation
Only the stated page rectangle is retained; original labels, plotted content and colors are unchanged. Non-figure page-link annotations are removed. Native PDF content is preserved, with a 240-dpi PNG for web display.
Evidence limit
This is an example response, not independent ground truth or the Table1 comparison. Its own output reports a power-balance issue; an answer still needs engineering review.
Original source / page
https://arxiv.org/abs/2509.02494v1
Original figure / image URL
https://arxiv.org/pdf/2509.02494v1

Visual credit: expanded_power_gridmind.png

Caption
GridMind 2026 Table 1: full reruns in fresh workspaces versus saved-result reuse, on a synthetic 2,000-bus hurricane-risk study.
Credit
Original chart from cited author-reported session/ablation measurements
Used on
53. GridMind results: fewer stages and less time are different
Asset type
Original quantitative teaching figure
Provenance note
GridMind weather-risk workflow, same seven-query session on a synthetic Texas 2000-bus case. Full-rerun ablation executes 35 stages and takes 1432.6 s; content-addressed reuse executes 14 stages and takes 677.1 s. The stage-count reduction is 60%; the calculated wall-time speedup is 2.1158× (2.12× rounded), corresponding to 52.7363% less wall time. These are different metrics, not two descriptions of the same percentage. No repeated-session uncertainty estimate is reported.
Evidence limit
Same reported query sequence; full-rerun arm uses fresh namespaces. Single simulation session, surrogate screening and reduced-order cascade models. No live utility operation or independently repeated performance claim.
Original source / page
https://openreview.net/pdf/5ad1c2afd7298f9e6ce3bc25f18e7165555e1ffd.pdf
Original figure / image URL
images/expanded_power_gridmind.png

Visual credit: real_paper_dynamicbench_overview.png

Caption
Original benchmark structure: input package, agentic system, tool calls and evaluation metrics.
Credit
Qian Zhang, Andrea Pomarico, Costas Mylonas, Magda Foti, Alberto Berizzi, Le Xie; PowerAgentBench-Dyn: A Benchmark for Agentic AI in Power System Dynamic Studies; Figure 1, PDF page 2
Used on
56. PowerAgentBench-Dyn allows only specific model changes
Asset type
Original published figure or result panel, native PDF crop
Transformation
Only the stated page rectangle is retained; original labels, plotted content and colors are unchanged. Non-figure page-link annotations are removed. Native PDF content is preserved, with a 240-dpi PNG for web display.
Evidence limit
Benchmark architecture, not the full PowerAgent roadmap or an operating utility deployment.
Original source / page
https://arxiv.org/abs/2606.20401v1
Original figure / image URL
https://arxiv.org/pdf/2606.20401v1

Visual credit: expanded_power_dynamic.png

Caption
PowerAgentBench-Dyn Table II: model-quality task. Other tasks in this benchmark use different models and checks.
Credit
Original chart from cited author-reported research results
Used on
57. PowerAgentBench-Dyn results: passed sessions and updates
Asset type
Original quantitative teaching figure
Provenance note
PowerAgentBench-Dyn DMQ case: ten independent sessions per model on one modified WECC solar PV dynamic model. Full-suite compliant success is 10/10 for Opus 4.8, 10/10 for Sonnet 4.6, and 9/10 for Haiku 4.5. Median updates are 1, 1 and 2. Eight quality tests and a maximum of five parameter updates define this case. No confidence interval or fleet-level performance is claimed.
Evidence limit
Author-reported counts and medians reproduced, not independently rerun. One fixed corrupted model; no human or deterministic search comparator.
Original source / page
https://arxiv.org/abs/2606.20401v1
Original figure / image URL
images/expanded_power_dynamic.png

Visual credit: real_paper_veragrid_agent_workflow.png

Caption
Original VeraGrid-Agent workflow: question, LLM controller, input file, AC-OPF workspace and selected result reads.
Credit
Shivanshu Tripathi, Hamed Mohsenian-Rad, Maziar Raissi; VeraGrid-Agent: Tool-Augmented LLMs for Distribution Optimal Power Flow at the Grid Edge; Figure 2, PDF page 2
Used on
59. VeraGrid-Agent uses solver results to answer a question
Asset type
Original published figure or result panel, native PDF crop
Transformation
Only the stated page rectangle is retained; original labels, plotted content and colors are unchanged. Non-figure page-link annotations are removed. Native PDF content is preserved, with a 240-dpi PNG for web display.
Evidence limit
A research question-answering workflow on one feeder. It is not a general feeder-control policy.
Original source / page
https://arxiv.org/pdf/2607.25155v1
Original figure / image URL
https://arxiv.org/pdf/2607.25155v1

Visual credit: power_study_veragrid.png

Caption
VeraGrid-Agent Table I: selected comparisons using the same model. Missing or badly formatted answers count as incorrect.
Credit
Original plot redrawn from cited primary-paper numerical results; no experiments rerun
Used on
60. VeraGrid-Agent results: the same models with and without tools
Asset type
Original quantitative teaching figure
Provenance note
Three explicitly selected same-model comparisons from VeraGrid-MCQ-150: GPT-5.2 64/150 without tools versus 149/150 with the solver-agent; Claude Sonnet 4.5 71/150 versus 149/150; Gemini 3 Flash 74/150 versus 150/150. Questions use one balanced 8-bus feeder and a VeraGrid reference solve.
Evidence limit
First three Table I rows selected for legibility, not ranked. Template-derived MCQs share one solved feeder. Repeated-run uncertainty is unspecified; results do not estimate field-control reliability or arbitrary-feeder generalization.
Original source / page
https://arxiv.org/pdf/2607.25155v1
Original figure / image URL
images/power_study_veragrid.png

Visual credit: real_paper_agentigrid_architecture.png

Caption
Original AgentiGrid data flow: loop controller, analysis-controller interface and ExaGO application.
Credit
Samim Konjicija, Slaven Peles; Integrating Agentic Artificial Intelligence with High-Performance Computing for Grid Planning; Figure 1, PDF page 2
Used on
61. AgentiGrid: what does one reported run mean?
Asset type
Original published figure or result panel, native PDF crop
Transformation
Only the stated page rectangle is retained; original labels, plotted content and colors are unchanged. Non-figure page-link annotations are removed. Native PDF content is preserved, with a 240-dpi PNG for web display.
Evidence limit
Prototype architecture for modeled studies. Twelve reported searches are different backend/application settings, not twelve repeats of one task.
Original source / page
https://arxiv.org/abs/2609.04544v1
Original figure / image URL
https://arxiv.org/pdf/2609.04544v1

Visual credit: rigor_r01_coverage.png

Caption
Counts read from R01 Table III and section IV. Each outage label is a pair of bus numbers in this test grid.
Credit
Original plot derived from cited paper data; see caption for interpretation
Used on
63. Successful tool calls can still miss a critical outage
Asset type
Original quantitative teaching figure
Provenance note
Derived from R01 Table III: the four reported backend searches on one synthetic ACTIVSg200 SCOPFLOW case all found economic-critical contingency 187–189 (+21.6% cost), three identified infeasible outage 29–30, and none identified infeasible outage 15–16. These are fixed backend-run counts, not repeated IID trials, estimated probabilities, or comprehensive field reliability. No confidence interval is justified by these four heterogeneous searches.
Evidence limit
One synthetic case, four different backend searches, reported search behavior only.
Original source / page
https://arxiv.org/pdf/2609.04544v1
Original figure / image URL
images/rigor_r01_coverage.png

Visual credit: rigor_r02_capability.png

Caption
R02 Table II. L1: place a 100 MW + 10 Mvar load. L2: check reactive-power needs for 100 MW + 50 Mvar generation.
Credit
Original plot derived from cited paper data; see caption for interpretation
Used on
64. Show both trial counts and uncertainty
Asset type
Original quantitative teaching figure
Provenance note
Derived from R02 Table II, PDF page 4, tested development snapshot. Bars reproduce the eight reported 100/1000-bus count cells. Whiskers are calculated two-sided Wilson 95% intervals, not author-reported uncertainty. Their probabilistic interpretation requires comparable independent Bernoulli attempts, which the paper does not establish. The OpenAI table label is gpt 5.4 B. Larger grids and levels L3–L6 have no quantified result cells and are omitted, not encoded as zero. This is not field-deployment reliability or a current vendor ranking.
Evidence limit
Reported counts preserved. Wilson intervals conditional on independent comparable trials. Model/backend and task snapshot specific.
Original source / page
https://arxiv.org/pdf/2607.28822v1
Original figure / image URL
images/rigor_r02_capability.png

Visual credit: real_paper_powerbench_dataset.png

Caption
Original PowerBench overview linking structured asset state, telemetry, rules, records and queries.
Credit
Xijing Wang, Yinsheng Yao, Jinru Ding, Yidong Jiang, Ziwen Xu, Yiwen Jiang, Jie Xu, Dawei Cheng; PowerBench: A Benchmark for Agentic Retrieval and Reasoning in Power Systems; Figure 1, PDF page 4
Used on
65. PowerBench finds answers across linked records
Asset type
Original published figure or result panel, native PDF crop
Transformation
Only the stated page rectangle is retained; original labels, plotted content and colors are unchanged. Non-figure page-link annotations are removed. Native PDF content is preserved, with a 240-dpi PNG for web display.
Evidence limit
Synthetic benchmark data, not utility operational records, simulation results or a live grid database.
Original source / page
https://arxiv.org/abs/2609.34492v1
Original figure / image URL
https://arxiv.org/pdf/2609.34492v1

Visual credit: expanded_power_benchmark.png

Caption
PowerBench Table 4: selected models and the authors' 95% bootstrap intervals. The two metrics average results differently.
Credit
Original chart from cited author-reported scores and bootstrap intervals
Used on
66. PowerBench results: complete answers and partial credit
Asset type
Original quantitative teaching figure
Provenance note
Selected PowerBench Table 4 results. Macro Joint Accuracy equally weights the three task families, with 60 temporal, 90 contextual and 150 comparative questions; it is not a correct-count divided by 300. Overall Field Accuracy is the mean per-question fraction of correct fields. Whiskers are author-reported 95% bootstrap intervals, not newly calculated intervals. All three configurations used the same 300-question benchmark and 50-call/300-second retrieval-and-reasoning budgets.
Evidence limit
Different aggregation rules are retained. Partial field credit is not full-task success. Do not infer pairwise significance from marginal interval overlap. Benchmark results are synthetic and author-reported; public runnable release not verified.
Original source / page
https://arxiv.org/abs/2609.34492v1
Original figure / image URL
images/expanded_power_benchmark.png

Visual credit: real_paper_powerbench_tasks.png

Caption
Original PowerBench task examples. Each task group contributes equally to the reported average.
Credit
Xijing Wang, Yinsheng Yao, Jinru Ding, Yidong Jiang, Ziwen Xu, Yiwen Jiang, Jie Xu, Dawei Cheng; PowerBench: A Benchmark for Agentic Retrieval and Reasoning in Power Systems; Figure 2, PDF page 6
Used on
67. A strong average can hide a weak task group
Asset type
Original published figure or result panel, native PDF crop
Transformation
Only the stated page rectangle is retained; original labels, plotted content and colors are unchanged. Non-figure page-link annotations are removed. Native PDF content is preserved, with a 240-dpi PNG for web display.
Evidence limit
Illustrates three task families; the metric gives them equal weight even though their sizes are60,90 and150 questions.
Original source / page
https://arxiv.org/abs/2609.34492v1
Original figure / image URL
https://arxiv.org/pdf/2609.34492v1

Visual credit: rigor_reliability_chain.png

Caption
Illustrative calculation, not measured agent performance: P(all L steps succeed) = q^L.
Credit
Original illustrative calculation for this tutorial; not measured agent performance
Used on
69. More required steps can lower overall reliability
Asset type
Original quantitative teaching figure
Provenance note
Illustrative calculation: if every step has identical success probability q and steps are independent, all L required steps succeed with probability q^L. These are mathematical scenarios, not measured agent reliability. Correlated errors or unequal steps require a different model.
Evidence limit
Identical independent Bernoulli step outcomes. All required steps must succeed.
Original figure / image URL
images/rigor_reliability_chain.png

Visual credit: rigor_eval_outcomes.png

Caption
Invented example: 80 feasible and 20 infeasible cases. These counts explain metrics; they are not deployment results.
Credit
Original illustrative calculation for this tutorial; not measured agent performance
Used on
70. The answer changes when we change what we count
Asset type
Original quantitative teaching figure
Provenance note
Hypothetical evaluation of 100 cases, not measured system performance. Of 80 truly feasible cases, 70 are accepted and 10 rejected. Of 20 truly infeasible cases, 2 are accepted and 18 rejected. False approval among infeasible cases is 2/20=10%, while the pooled fraction is 2/100=2%. Accepted-case precision is 70/72=97.22%. The denominator answers a different question in each metric.
Evidence limit
Illustrative labels and outcomes, fixed failure definition. No measured agent data.
Original figure / image URL
images/rigor_eval_outcomes.png

Visual credit: rigor_zero_failure_bound.png

Caption
Calculated binomial bound. It needs a fixed failure definition, independent trials and stable test conditions.
Credit
Original illustrative calculation for this tutorial; not measured agent performance
Used on
71. Zero test failures do not prove perfect performance
Asset type
Original quantitative teaching figure
Provenance note
Exact one-sided 95% binomial upper confidence bound after zero failures: p_upper = 1 - 0.05^(1/n). At n=30 the bound is 9.503%, at n=60 it is 4.8703%, at n=300 it is 0.9936%, and at n=3000 it is 0.09981%. It describes uncertainty under independent, identically distributed trials with a fixed failure definition and stable operating conditions. It does not establish reliability under a new topology or a new task distribution.
Evidence limit
IID binomial trials, zero observed failures, prespecified one-sided 95% confidence procedure. This bound is calculated, not author-reported.
Original figure / image URL
images/rigor_zero_failure_bound.png

Visual credit: real_framework_mcp_architecture.png

Caption
Original MCP architecture. Clients connect to servers; the engineering tool performs the calculation.
Credit
Source: Model Context Protocol, official documentation revision July 28, 2026. Screenshot captured October 18, 2026.
Used on
73. MCP connects a workflow to its tools
Transformation
Native browser screenshot of the original rendered source diagram only. Page navigation and surrounding prose excluded; diagram content unchanged.
Evidence limit
A protocol architecture does not prove numerical correctness or that a utility uses this implementation.
License or permission status
Publisher attribution supplied. No separate permission or unrestricted reuse license claimed.
Original source / page
https://modelcontextprotocol.io/docs/2026-07-28/learn/architecture
Original figure / image URL
https://modelcontextprotocol.io/docs/2026-07-28/learn/architecture

Visual credit: real_ops_frequency.png

Caption
Real CAISO oscillation-monitoring screens. The operator must still interpret the result and decide what to do.
Credit
California ISO, John Phipps, Control center modernization is enhancing reliability, 10 September 2025.
Used on
75. Measure the whole job, including human review
Asset type
Original operational software screenshot
Transformation
Original image; format conversion only where required.
Evidence limit
Context photograph or conventional operator tool; no agent performance or control authority is implied.
Original source / page
https://www.caiso.com/about/news/energy-matters-blog/control-center-modernization-is-enhancing-reliability
Original figure / image URL
https://www.caiso.com/content/images/blog/blog-graphics/controlcentermodernization-rtdms.webp

Visual credit: real_ops_training.png

Caption
California ISO’s operator-training facility. Photo: California ISO, 2025.
Credit
California ISO, Tricia Johnstone, ISO now has expanded training and emergency operations facilities, 20 February 2025.
Used on
76. Test whether the assistant helps on comparable tasks
Asset type
Original utility training-facility photograph
Transformation
Original image; format conversion only where required.
Evidence limit
Context photograph or conventional operator tool; no agent performance or control authority is implied.
Original source / page
https://www.caiso.com/about/news/energy-matters-blog/iso-now-has-expanded-training-and-emergency-operations-facilities
Original figure / image URL
https://www.caiso.com/content/images/blog/blog-graphics/training2.webp

Visual credit: ornl_goal.jpg

Caption
ORNL GOAL research laboratory, shown for context. This is not a photograph of LLNL's testbed.
Credit
Oak Ridge National Laboratory, U.S. Department of Energy
Used on
77. Power-sector testbeds provide places to evaluate agents
Asset type
Photograph
Transformation
None
Evidence limit
The source describes a research test bed and replicated utility command center. Do not label this as a production utility deployment.
License or permission status
Official ORNL laboratory photograph. No explicit image-specific reuse license found. ORNL is managed by UT-Battelle for DOE, so public-domain status is not assumed. Permission not independently obtained.
Original source / page
https://www.ornl.gov/grid-operations-and-analytics-laboratory-goal/
Original figure / image URL
https://www.ornl.gov/wp-content/uploads/2026/06/2021-p07416_1-970x647.jpg

Visual credit: gridguard_research_feeder.png

Caption
Existing workshop figure: synthetic SMART-DS Greensboro feeder udt 8894. This is not a live utility circuit.
Credit
Existing GridGuard workshop; SMART-DS synthetic Greensboro feeder udt8894
Used on
78. GridGuard: what should this prototype prove?
Asset type
Existing scientific model plot
Transformation
Copied without alteration
Provenance note
Local existing workshop figure: Agentic_Grid_LaTeX_Source/figures/feeder.png
Evidence limit
Synthetic research feeder, not a live utility circuit or measured deployment outcome.
License or permission status
Existing user-workspace teaching figure. Source and model identity preserved.
Original source / page
research/Follow_on_workshop_source.txt

Original diagram: A generative-model call produces content

Credit
Original editable TikZ teaching diagram; source and assumptions in caption
Used on
3. Generative AI creates content
Caption
Original example of a single generative-model call: a prompt and selected records produce a draft report for review. This illustrates one use of generative AI, not a universal boundary around the technology; the same kind of model can also participate in an agent.
Format
Native TikZ in the LaTeX edition; vector SVG in the web edition
SVG diagram
diagrams/va_generation_call.svg
Editable TikZ source
diagrams/va_generation_call.tex
Concept source [DEF03]
https://www.anthropic.com/engineering/building-effective-agents

Original diagram: A valid number can still violate the tool contract

Credit
Original worked teaching illustration
Used on
23. Unit checks catch a kV versus p.u. mistake
Caption
Illustrative typed contract. A solver expecting per-unit voltage and topology v43 rejects 12.47 kV and v42. A request with voltage 1.00 pu on base 12.47 kV and topology v43 can pass those checks. Unit conversion requires an explicit, matching voltage base and convention. Topology consistency is a separate requirement.
Format
Native TikZ in the LaTeX edition; vector SVG in the web edition
SVG diagram
diagrams/typed_contract.svg
Editable TikZ source
diagrams/typed_contract.tex

Original diagram: A topology event invalidates the old study context

Credit
Original worked teaching illustration
Used on
24. A study must match the current grid connections
Caption
Illustrative two-feeder topology, not an electrical solution. Switch S is closed in snapshot v42 at 14:35, opens at 15:02, and produces state v43. The 14:40 study used v42. At 15:05, the operator must revalidate its applicability to v43 before relying on it.
Format
Native TikZ in the LaTeX edition; vector SVG in the web edition
SVG diagram
diagrams/snapshot_consistency.svg
Editable TikZ source
diagrams/snapshot_consistency.tex

Original diagram: Exercise 1 answer: C passes the first checks

Credit
Original editable TikZ worked answer for this tutorial
Used on
27. Exercise 1 answer: C passes the first checks
Caption
The 60-second limit and packet labels belong to this teaching exercise.
Format
Native TikZ in the LaTeX edition; vector SVG in the web edition
SVG diagram
diagrams/exercise_packet_solution.svg
Editable TikZ source
diagrams/exercise_packet_solution.tex

Original diagram: Three graphs with different meanings

Credit
Original editable TikZ teaching diagram; source and assumptions in caption
Used on
34. Three kinds of graph have different jobs
Caption
Original schematics: undirected physical equipment connections, directed calculation prerequisites with fork/join dependencies, and runtime execution routes with a retry cycle. The physical network and all labels are illustrative; the diagrams are not a minimal DAG or a measured paper trace.
Format
Native TikZ in the LaTeX edition; vector SVG in the web edition
SVG diagram
diagrams/va_three_graphs.svg
Editable TikZ source
diagrams/va_three_graphs.tex
Concept source [LG01]
https://docs.langchain.com/oss/python/langgraph/graph-api
Concept source [D01]
https://www.manuscriptlink.com/society/kics/media?key=kics%2Fconference%2Ficaiic2025%2Fpresentation%2F1571098105.pdf

Original diagram: A fault-record DAG follows the data prerequisites

Credit
Original editable TikZ teaching diagram; see caption for source and assumptions
Used on
35. A DAG shows which result must come first
Caption
Original teaching adaptation of DrAgent's domain tools, not its measured execution trace. A frozen COMTRADE record R17 feeds independent find_fault_inception and find_max_fault_current computations (abbreviated in the drawing). The inception sample and sampling rate feed a time conversion; the report waits for both numerical results.
Format
Native TikZ in the LaTeX edition; vector SVG in the web edition
SVG diagram
diagrams/power_fault_dag.svg
Editable TikZ source
diagrams/power_fault_dag.tex
Concept source [D01]
https://www.manuscriptlink.com/society/kics/media?key=kics%2Fconference%2Ficaiic2025%2Fpresentation%2F1571098105.pdf

Original diagram: Check the prerequisite before executing a state-dependent tool

Credit
Original native TikZ educational diagram; source and assumptions in caption
Used on
42. PowerDAG: solve the grid before checking its voltage
Caption
Original teaching adaptation of PowerDAG v4, Section IV-C, PDF p.5. A planner proposes a voltage-violation check. A prerequisite gate blocks execution and returns the missing solve to the planner when the required solved network state is unavailable; otherwise the check proceeds. This is neither exact source code nor the complete/minimal workflow DAG.
Format
Native TikZ in the LaTeX edition; vector SVG in the web edition
SVG diagram
diagrams/revision_powerdag_gate.svg
Editable TikZ source
diagrams/revision_powerdag_gate.tex
Concept source [PD01]
https://arxiv.org/pdf/2603.17418v4

Original diagram: A changed input determines the stages that rerun

Credit
Original editable TikZ teaching diagram; see caption for source and assumptions
Used on
52. GridMind: which changes need a new run?
Caption
Original affected-stage diagram derived from GridMind 2026 Table 1, PDF p.4. Changing cascade fan-out k from 50 to 200 reruns cascade only. Shifting forecast initialization by six hours reruns all five stages: line-wind, outages, ACOPF, contingency and cascade. These are reported dependency effects, not a newly executed workflow.
Format
Native TikZ in the LaTeX edition; vector SVG in the web edition
SVG diagram
diagrams/power_gridmind_invalidation.svg
Editable TikZ source
diagrams/power_gridmind_invalidation.tex
Concept source [P03]
https://openreview.net/forum?id=Cs5TBzzNym

Original diagram: Exercise 2 answer: rerun what changed

Credit
Original editable TikZ worked answer for this tutorial
Used on
55. Exercise 2 answer: rerun what changed
Caption
Reuse follows the cited workflow’s input dependencies; another implementation can differ.
Format
Native TikZ in the LaTeX edition; vector SVG in the web edition
SVG diagram
diagrams/exercise_reuse_solution.svg
Editable TikZ source
diagrams/exercise_reuse_solution.tex
Concept source [P03]
https://openreview.net/forum?id=Cs5TBzzNym
Concept source [LG02]
https://docs.langchain.com/oss/python/langgraph/persistence

Original diagram: Retrieved instructions cannot acquire command authority

Credit
Original worked teaching illustration
Used on
74. A retrieved document cannot give itself permission
Caption
Illustrative prompt-injection mechanism. A retrieved document contains an instruction to ignore limits and execute. Its contents remain data for retrieval and proposal formation. An independent policy and permission gate stands before the command service. The direct path from document text to command execution is blocked.
Format
Native TikZ in the LaTeX edition; vector SVG in the web edition
SVG diagram
diagrams/trust_boundary.svg
Editable TikZ source
diagrams/trust_boundary.tex
Concept source [T09]
https://modelcontextprotocol.io/docs/2026-07-28/tutorials/security/security_best_practices