Claude Managed Agents get hard spend caps and mid-turn advisors
One parameter now halts a runaway agent at a set dollar cost; three more add advisor models, repo skill-loading, and region-pinned inference.
Copy markdown
Cap a session's spend before it runs away
Managed Agents now accept a hard budget at public list rates, and the session pauses with a budget_reached stop reason instead of draining your balance. Raise or remove the cap to resume — the same budget parameter also works on deployments.
An advisor model that coaches mid-turn
Add an advisor entry ({"type":"advisor"}) to your multiagent roster and Claude consults it mid-turn for strategic guidance. The catch: the advisor model must be at least as capable as the agent it's steering.
Your repo's skills load themselves
Point a session at a GitHub repo and it auto-discovers and loads everything in .claude/skills at session start — no manual registration. Commit a skill once, and every agent run picks it up.
Pin where your inference runs
Set inference_geo on the model object to keep a session's inference in a chosen region, overridable per session. Custom pricing applies per geo — the switch you need when data residency is a hard requirement.