Human-in-the-Loop and Approvals

Approvals

Part of: AI Agents & Tool Use

In 2024 an experimental coding agent, told to "clean up the project," confidently ran a command that wiped a developer's uncommitted work. The agent did exactly what it decided was right. The problem was that nobody stood between its decision and the irreversible action. The fix is not a smarter model, it is an approval gate : a checkpoint where a human says yes before the agent does anything it cannot take back. What it is Human-in-the-loop (HITL) means inserting a person's approval into the agent's action loop for high-stakes steps. The agent does not execute a gated action directly. It proposes the action, the runtime pauses , a human approves or rejects , and only an approved action runs. The key idea is sorting actions by risk. Reading a file, searching the web, doing math, low stakes, run freely. Sending money, deleting data, emailing customers, deploying code, high stakes, gate them. You are drawing a line between actions that are reversible and cheap and actions that are irreversible or expensive . How it works The runtime keeps a set of risky tools. Before executing any proposed action, it checks: is this tool risky, and if so, did a human approve it? A safe approval gate

Challenge: The Approval Gate