Claude's new 'auto' policy lets the server gate every agent tool call
Server-side rules now run, deny, or pause each call; a new CLI attaches your terminal to live sessions; and Gemini learns to skim video on demand.
Copy markdown
Run, deny, or pause — the server decides
The Claude Developer Platform added an `auto` permission policy: instead of blanket-allowing tool calls or approving each by hand, the server evaluates every agent and MCP tool call and runs it, denies it, or pauses for your approval. The `agent.tool_use` and `agent.mcp_tool_use` events now carry an `evaluation` field beside `evaluated_permission`, so you can see exactly why each call passed or got blocked.
Attach your terminal to a running agent
A new CLI command, `ant beta:sessions connect`, hooks your terminal onto a live Managed Agents session: follow it in real time, send it messages, and allow or deny the tool calls waiting on approval. Pass `--web` to serve the Claude Console session viewer locally. It turns a fire-and-hope agent into one you can steer mid-run.
Elsewhere: Gemini learns to skim video
Google's Gemini API shipped agentic video understanding: rather than ingesting a whole clip, the model navigates the timeline and pulls transcripts, frames, or audio tracks on demand — up to 88% fewer tokens on long-form content. Cheaper video Q&A and highlight-extraction agents are now practical to build.