# Human-in-the-Loop Design — AI Agents and Tool Use

Source: https://www.geekswithgeeks.com/en/ai-agents-mcp/eval-human-in-loop

> Place human review where it adds the most safety for the least friction.

## Right checkpoints

Approving every step makes people click "yes" without reading. Approving nothing is risky. Put human checkpoints at **irreversible or external actions**, when **confidence is low** or the goal is ambiguous, and at the **end** for important outputs. Make review easy: show the proposed action, the reason and a one-click edit or reject.

## Track approval rates

If people approve 99% of a prompt unchanged, it may be safe to automate or narrow. If they often reject it, fix the tool or prompt behind it. The data tells you where review is earning its keep.

**Quiz:** Why not require approval for every single step?

- [ ] Approvals are illegal
- [x] People start approving without reading
- [ ] Agents cannot be paused
- [ ] It makes tools faster

*Answer:* People start approving without reading. Approval fatigue turns a safety control into a rubber stamp.