Claude Managed Agents Get Hard Session Spend Caps, a Mid-Turn 'Advisor' Model Role, and Per-Agent Inference Geo Control
Anthropic's platform release notes for August 7 add three things aimed squarely at production agent fleets: session budgets (a hard cap priced at public list rates; a session that hits it pauses with a new `budget_reached` stop reason and resumes if the budget is raised or removed, with deployments propagating the same budget to every session they start), an `{"type": "advisor"}` entry in the multiagent roster letting the primary thread consult a model at least as capable as the agent itself for mid-turn strategic guidance, and an `inference_geo` field on the model object for data-residency control per agent or per session. The budget primitive lands the same week Databricks published its cost-control playbook — spend caps are becoming a first-class API concept, not a dashboard afterthought.
↳ Follow the thread