September 14, 2026: Anthropic added Claude Managed Agents auto permissions on September 10. The new policy lets the server assess agent and MCP tool calls, then execute them, block them or request approval. The official release notes also announce evaluation details in tool-use events, giving integration teams more information about permission decisions.

What changes for agent workflows

According to Anthropic’s permission-policy documentation, auto considers the tool, its inputs and session context. It must be enabled explicitly; no toolset adopts it by default. Application-executed custom tools remain outside these policies.

An allowed call executes immediately. A denied call returns an error to the agent while the session continues, and clients cannot override that denial. An undecided call pauses for approval. Teams requiring a person to review every call should use always_ask for that tool.

Policy updates affect newly created sessions; existing sessions retain their original tool configuration. Events expose evaluated_permission and, in most cases, an evaluation object. Clients should tolerate unfamiliar evaluation types or reason codes.

Why this matters for QA engineers

This adds a decision point to test when an agent triages failures or invokes an MCP integration. A completed agent session alone does not prove that each intended tool actually ran. QATechTools recommends checking both the permission outcome and the resulting tool evidence.

For a focused pilot, exercise allow, deny and approval-required paths against disposable test resources. Verify that blocked actions leave resources unchanged, pending approvals appear clearly in the client, and subsequent agent output does not describe a blocked action as successful. These are suggested QA checks, not reported performance gains.

What to check before rollout

Create a fresh session after changing a policy and retain its configuration with the test results. Compare client behavior across the three outcomes, including error handling and recovery. Keep explicit human review on actions where your team requires it, then evaluate whether automatic decisions reduce interruptions without weakening those requirements.