F1 · Human in the loop

Family: Oversight and operations

Checkpoints where a human approves, corrects or takes over.

Principle

At chosen points, execution stops so a human can approve, correct or take over: blocking approval, notification, question or agent inbox. Approval can target one specific tool call rather than the whole output. On resume, some frameworks restart the node from the beginning: side effects placed before the interruption must therefore be idempotent.

The problem it solves

An agent acting alone can carry out an irreversible or high-stakes action before anyone can stop it.

When to use it

  • Actions are sensitive, irreversible or high-stakes (payment, large refund, cancellation), or failures exceed a threshold.
  • You aim for partial autonomy: the agent prepares, the human keeps the final decision.

When to avoid it

  • The agent’s reliability is measured and stable on the actions concerned: approval can then be relaxed gradually.
  • There is no channel to reach the human: the approval queue, notification and resume have to be built first.

Common pitfalls

The anti-patterns that misuse this pattern, best established first. Each sheet describes the symptoms, the public incidents and the remedies.

Oversight, operations and governance

All anti-patterns

Synonyms by vendor

  • Human-in-the-loop (Google Cloud, Google ADK, Microsoft)
  • Agent Inbox (LangChain)
  • Contact humans with tool calls (12-Factor Agents (facteur 7))

Neighbouring patterns

Sources

Content reviewed on 2026-10-02

Pattern catalogue