What changed
Anthropic’s 10 September platform release notes add an auto permission policy for Managed Agents. The server evaluates agent and MCP tool calls, then runs them, denies them or pauses for approval. Events expose information about the evaluation.
The same update adds a terminal connection command for following sessions, sending messages and responding to pending approvals.
Why it matters
An agent that can complete a task also needs a clear boundary around what it may change. Automatic evaluation does not remove the need to choose tools, access and approval requirements deliberately.
Our interpretation: start with an observable workflow, limited permissions and a small set of reversible actions. Expand the scope after reviewing actual behavior, rather than assuming a successful demonstration proves every action is safe.
Sources & further reading
Sources checked 2026-09-11. Product access and features may change. Practical guidance is our editorial interpretation.
