Module 2 · Composition and control · Lesson 5
Human approval for high-stakes actions
Decide which agent actions a human expert must approve, then build approval checkpoints that pause durably, show the right evidence, check who may approve, and survive multi-tenant deployment, in eve and the Claude Agent SDK.
By the end of this lesson you can
- Classify an agent's actions by reversibility and external effect, and choose the approval level for each
- Implement an eve tool with an approval request policy, a custom prompt and a response policy that restricts who may approve
- Implement the equivalent gate in the Claude Agent SDK with always_ask and canUseTool, and explain why allowedTools alone would bypass it
- Explain why approval is neither authorization nor deduplication, and add the missing controls
- Estimate the reviewer workload of different approval granularities and design against approval fatigue