Every agent and team is pinned to a model and a version, so teams manage exactly which behavior is live.
New behavior can be staged before it reaches production workflows.
Teams, sub-agents, and used-by relationships live in the catalog, not in tribal knowledge.
Every capability an agent can call is classified by access — read, write, or needs confirmation.
Each skill gets the tools its workflows support, nothing broader.
Sensitive actions stand out from the catalog and stay behind confirmation.
Review tool calls, status, tokens, and cost — with every run traceable to user, source, and outcome.
Operators can inspect what the agent did and why.
Spend, status, and ownership land in the log as runs happen, not after the fact.
Teams can manage which behavior is live.
Tool calls, outputs, and exceptions stay visible.
Sensitive actions remain governed by review paths.
