Fermion Fleet

Google All Things Agentic Hackathon — Track 3: The Fortified Enterprise Fleet · #AllThingsAgenticHackathon

A multi-agent fleet where the door is code, not a prompt — and you can watch an agent get stopped by it.

Open API docs Watch the demo video GitHub repo

What it is

A customer places an order and three roles handle it: triage → handler → auditor. The auditor's verdict is parsed by code, and the gate fails closed: every path that cannot read a boolean true (parse failure, missing field, the string "false", retry limit reached) resolves to stop. The ledger only locks the order on a real true.

On top of that sits a context engine (ctx-key) that decides what an agent may read at each step, what gets evicted when the window is full, and what is recalled when a later step needs it. The stack is Google's. The policy is ours.

What it runs on

Model	Gemini 3.5 Flash on Vertex AI (`global`)
Agent framework	Google ADK 2.7.1
Infrastructure	Cloud Run (dedicated service account, `roles/aiplatform.user` only)
Selection engine	`fleet/fermion_garden` — pure Python standard library

Try it

/docs — interactive API documentation (Swagger UI)
/health — liveness check, returns {"status":"ok"}
/openapi.json — the OpenAPI spec

Minimal example: create a session, then send one message to the fleet.

URL=https://fleet-843303850287.us-central1.run.app

# 1. Create a session (the first request may take ~15s: cold start)
SID=$(curl -s -X POST $URL/apps/fleet/users/u1/sessions \
      | python3 -c 'import sys,json; print(json.load(sys.stdin)["id"])')

# 2. Run one step (this calls the model for real)
curl -s -X POST $URL/run -H 'Content-Type: application/json' -d "{
  \"app_name\": \"fleet\", \"user_id\": \"u1\", \"session_id\": \"$SID\",
  \"new_message\": {\"role\": \"user\", \"parts\": [{\"text\": \"How much is the dark blue straw cup?\"}]}
}"

To see the gate hold and the recall happen, run the full three-step shift with 跑班.sh from the repository: URL=https://fleet-843303850287.us-central1.run.app ./跑班.sh

Honest limits

Sessions and the ledger live in process memory: if the service restarts they are gone, and the first request after idle can take ~15s (cold start). Firestore and Vertex AI Memory Bank were scoped and cut. This is one shift of three steps, run three times — it shows the mechanism works; it is not a benchmark.