Anthropic adds auto-evaluation permission mode and live session connect to Claude Managed Agents
2026-09-10
On September 10, Anthropic added an `auto` permission policy to Claude Managed Agents: the server now evaluates each agent or MCP tool call and runs it, denies it, or pauses for approval, reporting the outcome in `agent.tool_use`/`agent.mcp_tool_use` events. The `ant` CLI added `ant beta:sessions connect`, letting a user attach a terminal to a live Managed Agents session to follow it, send messages, and approve or deny pending tool calls, with a `--web` option to serve the Console's session viewer locally.
Significance 3: A first-party changelog feature giving operators server-side automatic evaluation plus live oversight of agent tool calls — meaningful agent-governance tooling, below landscape-changing threshold.
Implications · machine-drafted, not owner judgment
Server-side auto-evaluation plus live session attach narrows the gap between 'agent runs unsupervised' and 'human approves every call,' giving enterprises a middle tier of oversight that could make agentic deployments more palatable to risk-averse buyers. For the owner's org-readiness lens, this is a Capability-force tool: it doesn't fix organizational process, but it gives an accountable human a real intervention point Anthropic is betting enterprises will need.
- An enterprise case study citing auto-evaluation or live session connect as a purchase driver
- This evaluation field surfacing in a public post-incident review of an agent going off-script
Sources
- primary Claude Platform release notes retrieved 2026-09-12