Fermion Fleet: When the Door Is Code, Not a Prompt
This post was created for the Google All Things Agentic Hackathon. Autonomous agents can sound certain while still being wrong. That is not a prompt-quality problem; it is a boundary problem. Fermion Fleet is a small multi-agent system built around one constraint: an order must not lock because a model says it is ready. It may lock only after code can read a structured boolean approval. Demo:…
This project, Fermion Fleet, is a multi-agent system built around the constraint that an order must not lock until a model confirms it can read a structured boolean approval. The demo showcases this concept, with a handler drafting an order confirmation that may require revisions from an auditor. The ledger accepts the order only when a real boolean approval is provided.
It is essential to understand that the gate has a straightforward policy: the handler prepares a draft, the auditor checks for required fields, and the gate parses the auditor's output. Only a structured boolean approval can lock the order, thereby ensuring fail-closed security. The project also explores context management in long-running systems, employing a small context window and a recoverable pool.
Low-priority items are evicted and stored in a recoverable pool, which can be recalled when necessary. This design ensures that relevant information can be recalled later, even if it is not currently in the active window. The Fermion Fleet project is built using Google's runtime stack, including Gemini 3.5 Flash through Vertex AI, Google ADK for the agent structure, and Cloud Run for deployment.
The policy layer is the team's own creation, governing context eligibility, eviction, return, role actions, and handoffs to durable ledger entries. The architecture is kept simple, with context and the ledger running in process memory, without relying on managed services like Firestore. To run the system yourself, clone the repository and execute the provided script.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.