Human-in-the-Loop (HITL)

Human-in-the-Loop (HITL) is a design principle for AI systems where human judgment is required at critical decision points before the system can proceed with consequential actions. Rather than granting full autonomy, HITL workflows insert checkpoints where a person reviews, approves, modifies, or rejects the AI's proposed next step.

Key characteristics of HITL workflows include:

1. Approval Gates: The system pauses at predefined points, such as before deploying code, sending communications, or modifying production data, and waits for explicit human confirmation.

2. Graduated Autonomy: Low-risk actions (reading files, running tests) can proceed automatically, while high-risk actions (deleting resources, publishing content) require human sign-off, creating a tiered permission model.

3. Feedback Integration: Human corrections during review are fed back to the agent, improving the quality of subsequent steps within the same session and informing future system improvements.

4. Audit Trail: HITL systems naturally produce a record of what was proposed, what was approved, and what was changed, which is essential for compliance and debugging.

HITL is considered a best practice for production AI deployments and is a core feature of tools like Claude Code, GitHub Copilot, and enterprise automation platforms.

In the Agentic Development Handbook, HITL is the enforcement mechanism of the Gate-Based Governance pillar. Mandatory human gates are required for security-critical changes (authentication flows, encryption, access control), architectural decisions that cross module boundaries, and high-blast-radius deployments that affect production data or user-facing services. The Eval Harness runs automated checks first — linting, type checking, test suites, and golden-sample comparisons — and only escalates to a human reviewer when automated validation is insufficient.

When an agent encounters an obstacle it cannot resolve, the handbook defines a Four-Phase Escalation Ladder: (1) automated gates run and fail, (2) the agent retries with adjusted context, (3) a Blocker Flag is raised for human intervention, and (4) if the task remains stuck, a Rescue Mission is initiated where a human takes over to unblock the work and document the resolution for future agent runs. This structured escalation ensures that human attention is reserved for genuinely ambiguous or high-stakes decisions rather than routine approvals.