Gate: human-in-the-loop with custom tools

Why it matters

Human review only works if the agent stops at the right moment and waits. This exercise turns that pause into a mechanism rather than an instruction.

What to practise

In Claude Managed Agents, a beta API, declare decide and escalate as "type": "custom" tools beside the built-in toolset, and run twelve receipts against an expense policy. Each call pauses the session with an agent.custom_tool_use event and a requires_action stop reason until your code posts a user.custom_tool_result. Count how many receipts land in each lane. An agent that escalates everything exhausts reviewers; one that escalates nothing is dangerous. Read the production pattern too: a webhook on session.status_idled replaces the open stream while a person decides.

How it fits

Choosing which receipts are clear-cut and which escalate is 5.3.1’s timing decision. The escalated call is the item a 5.3.2 review queue holds. Here the model decides when to escalate, so compare that with 5.3.3’s rule that mandatory policy triggers outrank the model’s own confidence.