UI - https://github.com/enturesting/murmuration/tree/dev-cs/murmurationB

Backend Agentic Workflow - https://github.com/enturesting/murmuration/tree/dev/agentic_workflow

Grid-Aware Compute Agents

Two AI agents — one representing a grid operator, one representing a fleet of data centers — that negotiate over where and when to run GPU workloads. The compute side bids on power; the grid side surfaces conditions. Together they route compute toward whichever region is cleanest, cheapest, and least stressed, without violating SLAs.

This is a hackathon project. The agents are real (Claude via the Anthropic API). The grid is synthetic but calibrated against real PJM and ERCOT operating ranges, and the schema mirrors what the agents would consume in production from public data feeds.

The problem

Data centers are on track to be roughly 10% of US electricity demand within a few years. That load is concentrated geographically — Northern Virginia alone hosts the densest cluster of data centers on Earth. Two coupled headaches result:

• Grid operators (ISOs) struggle to maintain reliability as data center load grows faster than transmission and generation can be added.
• Data center operators face rising electricity prices, growing carbon scrutiny, and limited tools for shifting load.

Today these two sides barely coordinate. Grid operators treat data centers as inflexible load. Data centers treat the grid as a passive utility. We propose they should talk — and that the conversation can be mediated by AI agents acting on behalf of each side.

What this project demonstrates

A working two-agent system in which:

• A grid-side agent ("ISO operator") publishes structured briefings about grid conditions across multiple zones — prices, carbon intensity, stress level, peak status, load vs. forecast.
• A compute-side agent ("DC fleet ops") reads those briefings, evaluates candidate placements, and decides per job whether to schedule, reject, or escalate — playing by hard-bid rules (each job has a willingness-to-pay; the agent never schedules above bid unless the job's SLA forces it).
• Both agents reason in natural language but ground every quantitative claim in tool calls. No hallucinated MWs.
• The system is benchmarked against two non-LLM baselines (naive and heuristic) on the same simulated trace. The scorecard reports cost, carbon, reject rate, and zone placement distribution — apples-to-apples comparison.