AI Pattern · Safety

Human-in-the-Loop Pattern

The agent requests human approval before it acts.

For consequential actions, the agent pauses and surfaces what it intends to do, waiting for a human to approve, edit, or reject before executing. It keeps a person accountable for irreversible steps while still letting the agent do the planning and the safe work autonomously.

How it works

  1. 1Agent plans an action (send, delete, charge, deploy)
  2. 2Pause and surface the proposed action to a human
  3. 3Human approves, edits, or rejects it
  4. 4Execute only on approval, and log the decision

Strengths

  • Safety on irreversible, high-impact actions
  • Accountability and a clear audit trail
  • Builds trust in autonomous systems

Watch-outs

  • Slower, gated on human availability
  • Approval fatigue if overused
  • Needs a flow to surface decisions clearly

When to use it

Destructive or irreversible actions, and anything touching money, production, or external parties. Reserve auto-execution for safe, reversible steps.

Example prompt

Before any action that sends a message, spends money, or deletes data, output a PROPOSED ACTION block describing it and stop.
Proceed only after an explicit “approve”.