Anthropic has upgraded Claude Managed Agents with two controls aimed at making long-running agents safer and easier to operate.
A new auto permission policy evaluates every agent or MCP tool call and decides whether to run it, deny it, or pause for human approval. The accompanying event data records how each call was evaluated, giving developers a clearer audit trail.
Anthropic also added ant beta:sessions connect, which attaches a terminal to an active agent session. Operators can watch messages and tool calls live, send guidance, interrupt execution, and approve or reject sensitive actions. A local browser viewer provides similar controls and exposes every thread in multi-agent sessions.
Why it matters
The important shift is not greater autonomy alone—it is controllable autonomy. Production agents need policy enforcement, observability, and a reliable way for humans to intervene without terminating the session.
That management layer is becoming as important as the underlying model. It echoes the operational requirements we highlighted when OpenAI introduced its ONA cloud workspace: clear approval checkpoints, intervention controls, and records that explain what happened.
Anthropic is turning those requirements into platform primitives. Builders can let routine calls proceed automatically while reserving ambiguous or sensitive actions for review, then inspect and steer the agent while it continues working.
Claude Managed Agents remains in beta and requires Anthropic’s Managed Agents beta header.