How to Give a Coding Agent a Clear Stop Condition

A coding agent works best when it knows exactly when to stop. Without a stop condition it tends to keep polishing, expand the scope on its own, or pause to ask what to do next. Stating the finish line in your first message fixes all three. This guide shows how to write one.

What you need

• An agent session open in your project
• A single task you want done
• About 5 minutes

Step 1: Name the finish line first

Open your request with the condition that means done. The agent reads it as the goal and works backward from it instead of guessing how far to go.

Step 2: Make done verifiable

A good stop condition is something the agent can check itself. Tie it to a command that passes or fails so there is no ambiguity about whether the work is finished.

Finish when `npm test -- auth` passes and `npm run lint` reports no errors. Then summarize the diff and stop.

Step 3: Fence off the scope

List what is out of bounds. An explicit do-not list stops the agent from wandering into unrelated files while it chases the goal.

Result: the agent runs to a defined finish, verifies it, and stops, instead of looping or quietly growing the change.