Claude Platform moves tool-call approval into the API

Anthropic added an auto permission policy to Managed Agents that puts the server in charge of every tool call an agent makes. Each agent.tool_use or agent.mcp_tool_use event now carries an evaluation field and an evaluated_permission value. The call runs, gets denied, or pauses for approval, decided before your code ever sees it, per the Claude Platform release notes.

That's a real shift if you've built your own approval gate in front of agent tool calls. That logic used to sit in code you control, where you could log every decision and grep it later. Now part of that decision can happen server-side, on infrastructure you don't operate.

The same release adds ant beta:sessions connect, a terminal command that attaches to a running session so you can follow it, send messages, and approve or deny whatever's waiting on you. Add --web and it serves the Console viewer locally instead.

I build with Claude Code on my own projects, and a denial is easy to trust when you wrote the rule that produced it. It's a different kind of trust when the rule lives on a server you don't operate, and the only way to watch it happen is to open a terminal and attach.