{"slug": "fermion-fleet-when-the-door-is-code-not-a-prompt", "title": "Fermion Fleet: When the Door Is Code, Not a Prompt", "summary": "Fermion Fleet, a multi-agent system built for the Google All Things Agentic Hackathon, ensures orders lock only when code reads a structured boolean approval, not when a model claims readiness. The system uses a fail-closed gate where an auditor's output must be parseable, and a gardener component manages context by evicting and recalling items from a recoverable pool. Built on Google's runtime stack, the project keeps its policy layer separate and explicitly avoids claiming undeployed components like Firestore or Model Armor.", "body_md": "This post was created for the Google All Things Agentic Hackathon.\n\nAutonomous agents can sound certain while still being wrong. That is not a prompt-quality problem; it is a boundary problem.\n\n**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.\n\nIn the demo, a handler drafts an order confirmation. It misses required fields. An auditor sends it back. The handler rewrites. Only a valid review can release the order to the ledger.\n\nThe important part is not that the auditor is asked to be careful. The important part is that the ledger accepts only a real boolean approval from a parseable result.\n\nThe gate has a deliberately boring policy:\n\nThis is fail-closed by construction. Looks good, a persuasive explanation, an unexpected format, and a parser failure all resolve to **stop**.\n\nWe tested that boundary by breaking the auditor’s output format. The auditor could still identify a real hallucination in natural language. It sounded professional. But code could not read a structured approval, so the door stayed shut.\n\nThat is the project’s central idea: **the door is code, not a prompt.**\n\nThe other problem is context management. In a long-running system, context cannot expand forever. But forgetting should not mean permanently deleting facts that a later step may need.\n\nFermion Fleet uses a small context window and a recoverable pool:\n\ncustomer\n\n-> triage: writes the case file\n\n-> gardener: select / evict / recall\n\n-> gate: handler -> auditor -> parse, fail closed\n\n-> ledger: locks only on boolean true\n\nWhen the window is full, the gardener evicts low-priority items into a recoverable pool. That eviction is driven by pressure, not by a timer. Later, when a new step needs an earlier detail, the system scores and recalls that item.\n\nIn the recorded run, an after-sales commitment leaves the active window. A later customer question makes it relevant again; the system recalls it, and the handler can answer with details that were not present in the current conversation. Without recall, that answer would be impossible.\n\nThe runtime stack is Google’s:\n\nThe policy layer is ours:\n\nThat separation matters. A model can generate the next action; the system still needs explicit, inspectable rules for what that action is allowed to do.\n\nThis hackathon build keeps context and the ledger in process memory. A Cloud Run restart loses them. We deliberately do **not** claim Firestore, a managed memory service, Model Armor, or a background side-track as deployed components.\n\nThose are sensible next steps, but they are not part of this submission. The architecture and README draw only what runs now.\n\nThe repository contains reproducible instructions. The Cloud Run service is an API, rather than a browser UI. To run a complete shift against the public deployment:\n\ngit clone [https://github.com/wubian87/fermion-fleet](https://github.com/wubian87/fermion-fleet)\n\ncd fermion-fleet\n\nURL=[https://fleet-843303850287.us-central1.run.app](https://fleet-843303850287.us-central1.run.app) ./跑班.sh\n\nA cold start can take roughly 15 seconds.\n\nA reliable agent system should make its important no decisions boring and mechanical. The model can be creative inside the workflow; the boundary that grants permission should remain readable by code.\n\nThat is the experiment behind Fermion Fleet: make a rejection visible, make a retry auditable, and make the final lock depend on a value that cannot be talked into existence.``", "url": "https://wpnews.pro/news/fermion-fleet-when-the-door-is-code-not-a-prompt", "canonical_source": "https://dev.to/xiaojia-fermion/fermion-fleet-when-the-door-is-code-not-a-prompt-n2m", "published_at": "2026-08-27 11:55:44+00:00", "updated_at": "2026-08-27 12:18:59.519342+00:00", "lang": "en", "topics": ["ai-agents", "ai-safety", "developer-tools"], "entities": ["Google", "Fermion Fleet", "Cloud Run"], "alternates": {"html": "https://wpnews.pro/news/fermion-fleet-when-the-door-is-code-not-a-prompt", "markdown": "https://wpnews.pro/news/fermion-fleet-when-the-door-is-code-not-a-prompt.md", "text": "https://wpnews.pro/news/fermion-fleet-when-the-door-is-code-not-a-prompt.txt", "jsonld": "https://wpnews.pro/news/fermion-fleet-when-the-door-is-code-not-a-prompt.jsonld"}}