Agentic AI Systems in Electrical Power Systems Engineering: Current State-of-the-Art and Challenges

Abstract

Agentic AI systems have recently emerged as a critical and transformative approach in artificial intelligence, offering capabilities that extend far beyond traditional AI agents and contemporary generative AI models. This paper addresses the gap by providing a comprehensive review that establishes a precise definition and taxonomy for "agentic AI," aiming to distinguish it from previous paradigms. Concepts are introduced gradually, starting with a highlight of its diverse applications across engineering broadly, then four detailed state-of-the-art use-case applications within electrical power systems engineering — a domain where the impact of agentic AI systems is expected to be particularly significant — ranging from an advanced agentic framework for streamlining complex power system studies and benchmarking to a novel agentic AI system developed for survival analysis of dynamic pricing strategies in battery swapping stations. Finally, the paper discusses the unique failure modes of robust deployment of these autonomous systems and derives actionable recommendations for the design and implementation of safe, reliable, and accountable agentic AI systems.

1. Introduction

The field of artificial intelligence is undergoing a significant paradigm shift, moving from systems that passively generate content to more autonomous systems that can reason, react, play, and act to achieve complex goals. This evolution can be broadly traced through three key phases: traditional rule-based agent systems, then generative AI powered by large language models, and finally the recent emergence of agentic AI frameworks. While the generative model was a powerful but inherently "reactive" assistant — responding to a prompt to complete a discrete task without its own goals or the ability to verify its outputs and execute them in a real-world feedback loop — agentic AI represents an evolutionary leap integrating the reasoning capabilities of LLMs with the classical agent's ability to act.

Agentic systems are not merely content generators; they are goal-oriented systems that autonomously decompose a high-level objective into a sequence of executable sub-tasks — planning, memorizing, reflecting on performance, and interacting with external tools such as compilers, simulators, or web APIs to execute their plans and adapt to new information. Recent studies have demonstrated this closed-loop feedback mechanism that mimics the core workflow of a human engineer, in applications such as full ASIC generation and high-level synthesis. After a critical review of four prior surveys, this paper identifies a significant gap in the literature: the absence of a comprehensive study of state-of-the-art applications, opportunities, and challenges specific to engineering in general and electrical engineering in particular — a discipline holding one of the highest potentials for an agentic revolution, given its fit for clean energy transition and decarbonization, navigating regulatory complexities and interconnection bottlenecks, and assessing complex power-grid cybersecurity vulnerabilities.

2. Foundational theory of agentic AI and comparison with AI agents

An AI agent is defined as an autonomous or semi-autonomous computational entity that reasons over information and takes prompt-based actions to achieve generally specific objectives, with applications in customer service and industrial automation. Generative AI refers to systems designed to create new content (text, images, audio, video, code) that learn patterns from massive datasets, yet require a user prompt and do not operate with autonomous goals. Agentic AI merges these two ideas with a higher degree of autonomy and planning (via goal decomposition) along with adaptive, multi-modal reasoning. An agentic framework usually consists of an orchestrator, a delegator, and an ensemble of specialized AI agents: the orchestrator analyzes the framework's overall goal and decomposes it into smaller decompositions, the delegator passes them to the agents, and consolidates results once tasks are completed to move the framework toward its goal.

The paper summarizes comprehensive distinctions between the three paradigms: in initiation and control, the traditional agent is rule- or event-driven, the generative model is prompt-driven, and the agentic system is self-initiating and goal-directed; in autonomy level (low, medium, high with autonomous planning and execution), task scope (narrow and predefined, broad but single-pass, multi-step cross-domain workflows), and architecture (single loop, single-stage model call, hierarchical controller–executor pattern with explicit planning–acting–observing loops, shared memory, and dynamic tool access). It also reviews use cases of these paradigms in engineering and technology management — from a retrieval-augmented energy-market rules assistant, to self-evolving control algorithm design, to agentic frameworks in automotive engineering, 6G networks, vehicular parking, urban planning, and energy markets — highlighting the progression from reactive agents to goal-driven systems orchestrating multiple tools and sub-tasks in complex engineering workflows.

3. Domain-specific case studies in electrical power engineering

3.1 MCP-based agentic AI framework for streamlining power system studies and benchmarking: The first case study presents an agentic framework in which specialized AI agents access power system simulation tools via the Model Context Protocol (MCP) to benchmark simulation software (run times and computation accuracy) and then decompose a complex multi-objective study — power flow, contingency analysis, and short-circuit studies — invoking the best-suited agents to deliver the outcome in a full study report. Successes: successful benchmarking across pandapower, PSS®E, and PowerWorld with a full report generated. Limitations: no standardized official vendor-provided MCPs existed at the time of writing. Lessons learned: prompts must be explicit (e.g., naming the filesystem tool and file path) to avoid hitting LLM token limits, since the framework sometimes tried to read or rewrite the entire network file before passing it to the simulation software.

3.2 MCP-based agentic AI framework for enhancement of substation illumination studies: The second case study presents an agentic framework automating the entire illumination calculation and reporting workflow for an electrical substation under National Electrical Safety Code (NESC) requirements. The framework evaluates major equipment (breakers, transformers, switches), searches its internal database for matching 3D models, and when none is found invokes a "specialist 3D modeling agent" leveraging Blender via MCP to generate them from catalog 2D prints; it then drives the illumination software (e.g., AGI32) via MCP, iterating over fixture combinations, placements, and orientations in a grid search to meet the goals: ≤ 0.2 foot-candle outside the fence, ≥ 5 foot-candle near major equipment, and no over-illumination (≥ 30 foot-candle). The framework successfully produced an NESC-compliant report for a six-position ring-bus and a three-bay breaker-and-a-half station, but software memory constraints blocked a four-bay report on standard hardware, and human-in-the-loop (HITL) checks were essential to match generated models against vendor drawings.

3.3 Agentic AI framework to generate engineering bill of quantity (BoQ) from request-for-pricing (RFP) documents: Engineering consulting firms receive pricing requests from utilities and developers containing schedules, drawings, and bidding rules, whose analysis takes a full engineering team to produce a bottom-up bill-of-quantities estimate. The proposed agentic framework streamlines this via three agents: a contextual retrieval-augmented generation (RAG) agent extracting information from engineering standard design documents, a reference-project assessment agent providing quantities from similar past projects, and a master compiling agent assembling the BoQ in the required format. In a pilot, the framework successfully digested lightweight RFP documents and compiled an Excel BoQ, but accuracy remained a concern with estimates falling within ±70% of a fully human-compiled BoQ, indicating the need for further refinement, plus scalability challenges with large volumes of engineering standards and the growing importance of choosing the right chat model for cost and response time.

3.4 Agentic AI system for survival analysis of battery swapping station pricing strategies: The fourth case study presents an agentic framework identifying the most suitable pricing strategy for EV battery swapping stations using survival analysis, where the event is defined as the "first instance of monthly profits ≥ $150,000" within a 12-month follow-up, with sub-threshold observations treated as right-censored. The orchestrator/delegator invokes a pricing-data compiling agent and then a specialized survival-analysis subagent executing the full workflow in R Studio via MCP: Kaplan–Meier estimation, log-rank testing, and Cox proportional-hazards modeling with 95% confidence intervals and proportional-hazards assumption checks (adaptively switching to a stratified Cox model upon violation). The system successfully identified the most profitable strategy across three metropolitan regions, achieving an 18% improvement in customer retention, and can be extended to first-life EV battery degradation analysis and second-life grid-storage classification, operating in a continuous loop with periodic data refreshes and alerts on hazard-ratio drift.

4. Failure mode investigations and recommendations for safe and accountable agentic AI

4.1 Adversarial spread of false information among networked LLM agents and mitigation: Agentic systems face a far broader attack surface than traditional systems: a malicious third-party-hosted agent within the framework can embed manipulated information that spreads into other agents' models, poisoning collective memory (illustrated by the false claim that PV modules generate at night in a power-system optimization problem). To counter this, the paper proposes adapting a Zero Trust Framework (ZTF) on the principle of "never trust, always verify": strong cryptographic identity per agent (X.509 certificates or SPIFFE IDs) with mutual TLS, aggressive micro-segmentation with least-privilege access to contain the "blast radius," continuous information-integrity verification via data provenance, consensus-based verification among agents, and plausibility monitoring by an independent auditor agent, plus dynamic behavioral monitoring with anomaly detection and quarantine of deviating agents.

4.2 Cascading misinformation from LLM rewrites and mitigation strategy: As information propagates through a sequential network of agents, each slightly rewrites content and accuracy degrades progressively — an experiment on 30 National Electrical Code (NEC) queries scored with DeepEval/GEval (grounding, exactness, verifiability) showed essential content losing context after the second or third rewrite and going fully off-topic after the fifth. The proposed mitigation is an architecture passing data packets with two inner clusters: one cluster processed "as-is" by the agent's tools with reinterpretation or rewriting by its chat model strictly forbidden during receipt or transmission, and another cluster where reinterpretation is allowed, plus an interlinked "guard signal" for chain continuity whose cluster assignment is decided by a human-in-the-loop (or a parent-AI-agent-in-the-loop in future work). For retrieval and summarization tasks not requiring advanced reasoning, lightweight LLMs with high token throughput, low per-token cost, and large context windows are preferable.

4.3 Other practical considerations for building trustworthiness: First, human-in-the-loop (HITL) and human-on-the-loop (HOTL) control mechanisms: start with strict HITL controls, then transition to HOTL after satisfactory field testing, rotating oversight duties and monitoring engagement metrics to avoid alert fatigue — a HITL audit of the BoQ case study caught the framework arbitrarily generating MEPPI breaker pricing as a percentage escalation over the Siemens option because the database labeled the category as Mitsubishi breakers. Second, AI governance standards: Singapore's Model AI Governance Framework (agnostic to algorithm, technology, sector, and scale), corporate open-source toolkits (Microsoft's FairLearn and InterpretML; IBM's AI Fairness/Explainability 360), and IEEE standards under development (P2840 on licensing, P3396 on risk and safety, P7999 on ethics oversight). Third, audit trails via self-documenting agentic systems: log files storing algorithm setup, LLM and MCP versions, system configuration and random seeds, intermediate files, final solutions, and KPI progress — with immutability (cryptographic hashing or blockchains), escalating penalties for repeat offenders, and redundant logging characteristics.

5. Conclusions and future work

This paper has charted the emergence of agentic AI as a transformative new paradigm representing an evolutionary leap from traditional AI agents and reactive generative AI — with goal-oriented autonomous behavior, task decomposition and planning capabilities, and the ability to orchestrate constituent AI agents interacting with external tools, moving AI into the role of an active problem-solver. Its primary contribution was bridging the gap between high-level theory and practical high-stakes application in engineering: first by establishing a clear taxonomy distinguishing agentic AI's unique capabilities, then by presenting four state-of-the-art electrical engineering case studies — power system simulation software benchmarking and studies, automated substation illumination design, automated bill-of-quantity generation, and advanced survival analysis for EV pricing — alongside failure-mode investigations covering adversarial false-information spread and accuracy degradation through successive rewrites, with tangible mitigations including the Zero Trust Framework and information-clustering architecture, complemented by practical recommendations emphasizing human-in-the-loop, adherence to emerging standards, and self-documenting audit trails for better accountability.

Robust and scalable deployment hinges on several high-impact research trajectories: standardization of engineering tool integration via official stable vendor MCPs (e.g., PSS®E, PowerWorld, CDEGS) offered as standard on vendors' websites; development of a centralized open-access electrical engineering tool registry with MCP specifications for dynamic agent-side tool discovery; formal verification of inter-agent planning logic (model checking, temporal logic, theorem proving) to guarantee safety in critical domains; extension to parent-AI-agent-in-the-loop for countering misinformation propagation; and addressing latency and computational overhead by deploying latency-critical agents at the edge (substations and control centers) with compressed models and hardware accelerators plus decentralized event-driven coordination that does not interfere with existing protection and primary control layers. Focusing on these areas can pave the way for agentic AI systems that are not only powerful and autonomous but also secure, reliable, and fundamentally trustworthy.

Author contributions and data availability

Authors Soham Ghosh and Gaurav Mittal contributed conceptualization, data curation, formal analysis, methodology, software, validation, original-draft writing, and review and editing, with Ghosh additionally providing supervision and visualization. Funding: this research was conducted without any external financial support. Data availability: the original contributions presented in the study are included in the article/supplementary material; further inquiries can be directed to the corresponding author (sghosh27@ieee.org). The article was edited by Nikos Hatziargyriou (National Technical University of Athens) and reviewed by Iacovos Ioannou (University of Cyprus) and Ali Rushdi (King Abdulaziz University).

References

Full reference list (around 90 sources) includes foundational studies on generative agents (Park et al., 2023), agent architecture surveys (Wang et al., 2024), the four critically compared surveys (Acharya et al., 2025; Schneider, 2025; Ali et al., 2025; Bandi et al., 2025), PowerAgent for the agentic roadmap in power systems (Zhang and Xie, 2025), and Model Context Protocol literature (Anthropic, 2024a; Singh et al., 2025; Hou et al., 2025), available in the original PDF.