Registries for models, prompts and agents

A registry is the record of what is running, who owns it, what it is made of, what evidence justified shipping it, who approved it and when it must be reviewed or retired. For an LLM agent the "model" is one component among several: a release pins a model snapshot, prompt version, tool set and scopes, agent configuration and guardrail policy, and is identified by a content hash. Artefact registries (MLflow, Langfuse, cloud model registries) store the components. Discovery registries (AWS Agent Registry, Microsoft Agent 365, A2A catalogues) say what exists and how to call it. The governance inventory, run by the customer's model risk or AI governance team, links them: one entry per agent with a named owner, risk tier, eval and approval chain, and lifecycle dates. Your job is to make that entry accurate on day one and keep it accurate automatically after you leave.

• The release hash is the identity; the version number is for humans. Evals, approvals and traces bind to the hash.
• No passing eval for this exact release, no approval. Enforce it in code.
• The approver is never the builder or the owner.
• Rollback targets are part of the inventory. If their model snapshot retires, rollback fails when you need it.
• An agent with no active owner is an incident waiting for a date.