Human-in-the-loop★

Human-in-the-loop (HITL) is a collaborative development pattern where the developer actively monitors and guides the agent's work. Instead of walking away and checking the results later, you review the agent's proposed plans, inspect its file edits mid-flight, and redirect its approach in real time.

You operate in a Human-in-the-loop pattern whenever you:

• Use an IDE agent in "ask permissions" mode, reviewing each file write.
• Interject in the chat: "Stop, that's the wrong directory; check the components folder instead."
• Conduct Socratic planning interviews (Grilling) before letting the agent modify files.

When to Stay in the Loop

Choosing between staying in the loop and leaving the agent to run unattended (AFK) depends on the risk and reversibility of the task:

• High Risk (Stay in the Loop): Tricky database migrations, changes to authentication logic, or deploying to production. Catching a mistake early saves hours of debugging.
• Low Risk (Go AFK): Writing repetitive unit tests, migrating simple syntax styles, or compiling documentation. These tasks can be verified later using automated checks.

Low Risk / Easy to Verify  ──► Run AFK (Unattended)
High Risk / Irreversible   ──► Human-in-the-Loop (Active Steering)

The challenge of staying in the loop is attention overhead. If the agent requests permission for every trivial read operation, you will get approval fatigue. A well-designed harness uses strict Permission Modes to auto-approve safe reads while pausing for unsafe writes.

Field Applications

Developers toggle between in-the-loop and AFK states during a task:

1. Start (In-the-loop): Set up the design concept and verify the early files.
2. Middle (AFK): Once the pattern is clear, switch permissions to bypass and let the agent write the boilerplate.
3. End (In-the-loop): Review the diffs and run the test suite.