Claude Managed Agents: attach to a live run, approve each tool call
The ant CLI can now tail a server-hosted agent, send messages mid-run, and gate tool calls — or let a new auto policy make that call, log it, and move on.
Copy markdown
Attach your terminal to a live cloud run
`ant beta:sessions connect` hooks your terminal to a running Managed Agents session: follow it live, send messages, and allow or deny tool calls waiting on approval. Add `--web` to open the Console's session viewer locally instead. Server-hosted agents stop being fire-and-forget.
auto mode lets the server judge each tool call
A new `auto` permission policy evaluates every agent and MCP tool call server-side, then runs it, denies it, or pauses for you. Each decision lands in `agent.tool_use` and `agent.mcp_tool_use` events via new `evaluation` and `evaluated_permission` fields, so your app sees exactly why a call was held. Policy-driven autonomy without hand-rolling a gate.
What it unlocks: leave it running, step in when it matters
Together these turn a cloud agent into something you supervise like a pair session — leave a long job running, get held only on the calls auto won't clear, then hop into the terminal to approve or redirect. This weekend: point a Managed Agent at a big migration and actually trust it unattended.
Elsewhere: define your agent stack as code with ant apply
Anthropic's `ant apply` (ant CLI v1.30.0) lets you declare agents, environments, skills, memory stores, and deployments in repo files, with a `claude-lock.json` pinning versions. Your agent infra becomes reviewable and diffable like any other code.
Elsewhere: Gemini CLI 0.59 closes an SSRF hole
Gemini CLI 0.59.0 patches an SSRF in MCP OAuth metadata discovery and switches to fail-closed workspace trust — it won't load MCP servers in untrusted folders. Update if you run it anywhere near code you didn't write.