Overview
Agent Deployments is the system of record for every deployed AI agent. Every catalog entry, every match, and every cost estimate reads from this data.
Data model
A deployment combines master data (agent type, title, owning team), key figures (integration count, tool count, model/provider), commercial details (cost per 1k tokens or monthly budget, chargeback rate), and compliance-certification figures. Artifacts - model cards, evaluation reports, architecture diagrams - attach to the deployment, not to a catalog entry, and are therefore available to every document generated later.
Lifecycle
draft → active → paused → retired or superseded → archived
Archived deployments remain readable. There's no true deletion, because settlements and disclosures reference them.
Agent metadata feed
POST /import/feed accepts JSON or a ZIP of agent cards. In merge mode, only the deployments in the file are created or updated; replace additionally archives anything not in the file. For a nightly sync from an existing agent registry, replace is the right mode - but only once the export is verifiably complete.
Common Mistakes
- Business unit as free text.
business_unitshould reference a known cost-center ID; leading zeros in cost-center codes are otherwise lost. - Compliance certification added later. Optional at creation, but publishing any catalog entry fails without it.
- Missing model card. Without
category: "model_card", the template falls back to a placeholder graphic.