Murmuration

The grid and the AI compute fleet need to start talking. We built the protocol — and the agents that speak it.

Built in one day at the SCSP AI Hackathon · Electric Grid Optimization track · Washington DC · April 26, 2026.

Data centers are becoming one of the largest — and most concentrated — loads on the US grid, right as extreme weather makes the grid more fragile. Murmuration is a bilateral wire format that lets grid operators and flexible loads negotiate in real time: the grid publishes state and dispatch requests; data centers and home-battery fleets answer with standing flexibility envelopes, acks, and counter-offers. Two live Python agents (Claude-backed, with rule-based fallback) speak the protocol on each side, on top of load forecasting, anomaly detection, and a self-healing topology layer.

Run it (no API keys required)

Everything below runs fully offline — live data sources and Claude narration are optional enrichments via .env.example.

1. murmuration/ — the flagship (Python)

The end-to-end system demoed on stage: protocol bus with 12 Pydantic message types, GridAgent + ComputeAgent, 9 data centers + a 100-home virtual power plant, GBM load forecaster, z-score anomaly detector, networkx topology healer, tiered workload router, 9 scenarios, and a 3-tab UI (3D globe · flat map · story walkthrough) over FastAPI + WebSocket.

cd murmuration
python3 -m venv .venv          # Python 3.12+
.venv/bin/pip install -r requirements.txt
./run.sh
# open http://127.0.0.1:8765 — pick a scenario in the right rail (e.g. "Texas heat wave")

See murmuration/README.md for scenarios, endpoints, troubleshooting.

2. nictopia/ — the visual demo (React)

A standalone globe + flat-map replay of real archived grid incidents (Winter Storm Uri, the 2019 CAISO PSPS, the duck curve, Hurricane Helene) with primary-source citations and formula-derived dollar/carbon figures. All data ships as committed JSON caches — no backend.

cd nictopia
npm install
npm run dev
# open http://localhost:5173

3. agentic_workflow/ — the original prototype (Python)

Where it started: two Claude agents negotiating GPU-job placement over a synthetic 14-day, 4-zone grid replay with hard-bid economics (every job has a max willingness-to-pay; rejection is a correct outcome). Includes a reproducible merit-order data generator with day-ahead vs real-time prices and forecast-bust modeling.

cd agentic_workflow
python3 -m venv .venv && .venv/bin/pip install pandas numpy pyarrow anthropic python-dotenv
.venv/bin/python generate_grid.py   # writes data/*.parquet — no API needed
.venv/bin/python gridcache.py       # data-layer smoke test — no API needed
# .venv/bin/python runner.py        # full agent replay — needs ANTHROPIC_API_KEY, ~$2-4

How the pieces relate

The team built in parallel and merged deliberately: Rohan's agentic_workflow/ prototype proved the two-agent negotiation loop; a Copilot-built mock and Nick's nictopia/ explored how to make it legible to judges; Shashank's murmuration/ package unified the ideas into the live protocol system we staged. The written comparison that drove the final call is in docs/reference/shashank_branch_review.md.

Repo map

Path	What it is
murmuration/	Flagship: protocol + agents + forecasting + topology healer + 3-tab UI
nictopia/	React globe/flat-map demo over real archived incidents
agentic_workflow/	Original two-agent prototype with hard-bid job economics
MURMURATION.md	Pre-build design doc (historical)
PITCH.md	Pitch-deck rationale
docs/demo/	Stage materials: 4-slide deck, demo script, judge Q&A prep, presenter card
docs/reference/	Grid-physics honesty notes, implementation comparison, nictopia design

Data honesty

• CAISO data is fetched live (via gridstatus) when the network allows; ERCOT/PJM/MISO/NYISO/ ISO-NE/SPP use EIA-930 (optional key) and everything degrades to calibrated synthetic data.
• nictopia's incident replays are anchored to dated, committed snapshots with FERC/NERC/utility source URLs in each JSON.
• HIFLD transmission-line/substation overlays for the flagship flat map require a bulk download (paths in murmuration/README.md); without them the map simply omits those layers.
• The UIs load three.js/globe.gl/Reveal.js from public CDNs, so the browser needs internet even though the backends run offline.

Team

• Nick Allison (@enturesting) — nictopia globe demo, real-incident data anchoring, demo narrative + pitch materials, integration
• Teddy Allison — scenario decisioning, deep research and design thinking, real-incident data anchoring, demo narrative
• Rohan Dani (@RohanDani2) — agentic_workflow/ prototype: synthetic grid generator, hard-bid compute-agent economics
• Shashank Chikara — flagship murmuration/ package: protocol layer, live agents, anomaly detector, topology healer, 3-tab UI