Where do you make your agents stop and ask a human? Here's what's held up for me
retained text
I run a fair amount of real business work through a small team of agents (research, drafting, code changes, scheduled reports). The hard part hasn't been output quality. It's been deciding where a human has to say yes. A few rules that have held up: **Approve the plan, not every step.** Per-step approval sounds safe, but people stop reading after a few days. I have the lead agent write a plan first: owner per step, dependencies, and "done when" criteria. I approve that once. **A short always-stop list.** Send, pay, publish, delete, push to a shared branch, sign up for anything. Those raise an approval naming the exact target, even inside an approved plan. Reading, drafting, building on a branch and running tests just run. **Criteria a checker can verify.** "npm test passes" or "table shows all three plans" can be checked by a second agent. "Looks good" can't. And a passing check only covers what you wrote down, so it means "ready for review", not "done". **Batch the recurring stuff.** Scheduled jobs write drafts and end with one combined form (tick what goes out, list skips, free-text edits). Unticked items roll to the next day. Much easier to actually read than a stream of separate pings. **Narrow tools over instructions.** "Don't change anything" is a request; a read-only agent is a guarantee. Curious what others do: - Do you gate by action type, by cost, by confidence score, or something else? - Anyone found a way to keep per-item approvals meaningful long-term? Disclosure: I'm building an agent-team product, so I'm biased toward this setup. Not linking it here per the sub's rules.
document fetched 2026-11-12 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xt9gyy/where_do_you_make_your_agents_stop_and_ask_a