Skip to main content

Human-in-the-Loop

The agent pauses execution at defined checkpoints and surfaces a decision to a human — for approval, rejection, or editing — before continuing. This is essential for any agent that takes real-world actions where mistakes are costly or irreversible.

Structure​

Checkpoints can be triggered by action type (destructive operations), confidence level (agent is uncertain), cost threshold (expensive API calls), or policy rules (compliance requirements).

How It Works​

  1. Execute — agent works through the task normally
  2. Hit checkpoint — agent reaches a step requiring human approval
  3. Pause — execution halts and the agent surfaces the pending action with context
  4. Human decides — approve (continue as-is), reject (agent retries differently), or edit (modify the action and continue)
  5. Resume — execution continues from where it paused

Checkpoint triggers:

  • Action-based — always pause before delete, send, publish, deploy
  • Confidence-based — pause when the agent is uncertain
  • Cost-based — pause before expensive operations
  • Policy-based — compliance requires human sign-off

Key Characteristics​

  • Safety net — prevents costly mistakes from autonomous agents
  • Trust building — users gain confidence by seeing what the agent does before it does it
  • Latency cost — human review adds unpredictable delay
  • Not scalable — more checkpoints = more human time required
  • Context matters — the agent must present enough context for the human to decide

When to Use​

  • Agent takes real-world actions (send emails, modify databases, deploy code)
  • Mistakes are expensive or irreversible
  • Compliance requires human approval for certain operations
  • You're building trust with users and want to start with more oversight
  • The agent handles sensitive data or high-stakes decisions