Power Market Intelligence Agent

A governed, skill-based agentic AI reference implementation for evaluating power-market and AI data-center opportunities.

The project demonstrates how to translate domain knowledge into modular AI skills, orchestrate those skills, preserve evidence lineage, apply deterministic controls, generate structured recommendations, and evaluate the complete workflow without proprietary data or paid APIs.

Decision-support only. This repository is an educational demonstration. It does not provide investment, trading, engineering, regulatory, or legal advice.

Demonstrated business question

Given a proposed data-center or generation project, the agent answers:

• Is the site and market structurally attractive?
• What drives the local power price?
• Is the proposed generation economically viable?
• Which transmission, fuel, permitting, financing, and contract risks matter most?
• What hedge structure could stabilize project cash flows?
• What conditions must be resolved before an investment committee proceeds?

Agent architecture

flowchart LR
    U[Opportunity YAML] --> P[Planner]
    P --> S1[Site Viability Skill]
    P --> S2[Plant Economics Skill]
    P --> S3[Market Structure Skill]
    P --> S4[Price Formation Skill]
    P --> S5[Congestion & Reliability Skill]
    P --> S6[Contract & Hedge Skill]
    S1 --> R[Recommendation Engine]
    S2 --> R
    S3 --> R
    S4 --> R
    S5 --> R
    S6 --> R
    R --> G[Deterministic Guardrails]
    G --> H[Human Review Required]

Loading

Core implementation patterns

• Planner/orchestrator: Selects and runs domain skills.
• Skill library: Each skill has a stable contract and independent tests.
• Structured outputs: Pydantic models prevent free-form output drift.
• Evidence lineage: Every finding identifies the input fields supporting it.
• Deterministic guardrails: Critical conditions override optimistic aggregate scores.
• Execution trace: Records skill status, duration, and version.
• Human-in-the-loop: The agent never authorizes a transaction or trade.
• Offline-first demonstration: No API key is required.
• Golden-case evaluations: Known cases verify recommendation behavior.

Quick start

python -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate
pip install -e ".[dev]"

power-agent run \
  --input data/sample_data_center.yaml \
  --output output/report.json

power-agent evaluate --dataset evals/golden_cases.yaml
pytest -q

Expected sample result:

Recommendation: DEFER
Human approval required: True

Sample opportunity

The included case is a fictional 300 MW AI data-center campus paired with a natural-gas generation strategy. It uses synthetic assumptions for fuel price, heat rate, power price, transmission capacity, contract tenor, and development readiness.

No confidential or real transaction data is included.

Repository layout

src/power_agent/       Application, orchestration, models, and skills
data/                  Synthetic demonstration cases
evals/                 Golden evaluation cases
tests/                 Unit and integration tests
docs/                  Architecture, governance, threat model, and demo material
.github/workflows/      Continuous integration
output/                 Generated demonstration report

Domain model

The implementation encodes several fundamental relationships:

Fuel cost ($/MWh) = Heat rate (MMBtu/MWh) × Fuel price ($/MMBtu)

Variable generation cost = Fuel cost + Variable O&M + Carbon cost

Spark spread = Power price − Fuel cost

Delivered power balance = Generation + Imports + Storage discharge
                          − Exports − Storage charge − Losses

The simplified model intentionally does not replace a production-cost model, AC power-flow model, legal review, engineering study, or ISO market simulation.

Public demonstration sequence

1. Open data/sample_data_center.yaml and explain the assumptions.
2. Run the orchestrator.
3. Show the six skill outputs and evidence references.
4. Show the deterministic guardrail that changes the final recommendation.
5. Open output/report.json to demonstrate auditability.
6. Run the golden evaluation cases.
7. Open the Actions tab to show automated CI.

See docs/demo-script.md for a three-minute walkthrough.

Roadmap

• Entitlement-aware document retrieval
• Live ISO, EIA, weather, natural-gas, and transmission data adapters
• Scenario simulation and sensitivity analysis
• Production-cost optimization tool integration
• Geospatial site-screening tools
• PPA and heat-rate-call-option term analysis
• Model-provider gateway and prompt registry
• Reviewer feedback capture and decision-outcome evaluation

License

MIT. See LICENSE.