Agent OS
Certification and operations layer for an AI agent fleet. One law: no evals, no production. Registry, trace ledger, runtime broker, and loop guard — a governed stack for autonomous agents.
Agent OS is the governance infrastructure for running a fleet of autonomous AI agents safely in production. It enforces one hard law — no evals, no production — and provides the four planes that law runs on.
The registry holds agent cards: JSON documents that define each agent's job description, capability badge, and probation terms. The tower is the run ledger: an append-only trace of every execution with three honesty rules enforced in code, not policy. The broker is the runtime gate every agent run physically passes through — it checks the card, enforces grants (deny wins), writes the trace, and pages a human on violation. The harvest plane turns raw traces and prove records into proposed eval cases, which a human must promote before they count.
A loop guard sits alongside the broker with five local stuck-detectors and a clean-restart primitive — no model in the path, no LLM judgment call about whether the agent is spinning. The design principle throughout: authorization is structural, not conversational.