Anthropic has added a new set of operational controls to Claude Managed Agents.
The August 7 Claude Platform release notes list four changes. Sessions can now have a hard budget. Sessions can consult an advisor model. Agents or single sessions can control where model inference runs through inference_geo. Managed Agents sessions can also load skills from a GitHub repository when a mounted repository contains a root .claude/skills directory.
The budget behavior is the most concrete governance change. Anthropic says a session that reaches its budget pauses with the budget_reached stop reason instead of starting new model requests. Changing or removing the budget resumes it. Deployments can accept the same budget and apply it to each session they start.
Agent controls are moving from prompts into runtime policy
The release is small on the surface, but the direction is important. Managed agents are no longer just long-running prompts with tools. They need runtime policy that handles cost, supervision, locality, and reusable capability packages.
The advisor feature points at a multiagent pattern. A primary thread can consult a model at least as capable as the agent’s own model for strategic guidance. That keeps the primary agent in control of the session while giving teams a built-in route for review or planning help.
The inference_geo setting is the enterprise counterpart. It lets teams control where model inference runs for an agent or override that location for one session. Anthropic’s release note points readers to data residency details for available geographies and pricing, but the release note itself is enough to show the product direction: region is becoming a session-level agent property.
Repository-loaded skills round out the picture. A mounted repo can bring its own .claude/skills directory into a session, so an agent’s reusable operating knowledge can travel with the codebase.
