Human in the Loop
HITL done well is empowering, not theatre. A practical pattern catalogue.
HITL is not a checkbox
A lot of "human in the loop" systems end up as theatre. There's a human, technically, looking at a stream of approvals, clicking "yes" over and over. That isn't oversight. That's a rubber stamp.
Real HITL has three properties:
- The human has the information they need to decide.
- The human has the time and authority to say no.
- Saying no has a clean, low-friction path.
If any of those is missing, you don't have a human in the loop. You have a human on the hook.
A pattern catalogue
| Pattern | When to use it |
|---|---|
| Plan-then-approve | Multi-step agentic work (refactors, migrations). |
| Sample review | High-volume low-stakes decisions (tagging, triage). |
| Escalation on uncertainty | Model confidence below threshold → route to human. |
| Dual control | Irreversible or sensitive actions (prod deploys, customer comms). |
| Audit-only | Fully automated, but every action is reviewable after the fact. |
Pick the pattern that matches the blast radius of the decision, not the convenience of the team.
A heuristic
If a human couldn't reasonably review one action in this stream, you should not be reviewing all of them. Either reduce the stream, raise the threshold, or change the pattern.
