Run coding work in background agents

Coding requests are routed to a coding-focused background agent so the main conversation remains available and the work receives only the coding tool profile.

Write a bounded request

Include:

• the intended outcome;
• the selected repository or folder;
• files or services in scope;
• compatibility constraints;
• tests or acceptance criteria;
• actions that must not be taken.

Ask for diagnosis only when you do not want changes. Ask to implement or fix when writes are intended.

Follow progress

The main conversation acknowledges the delegated task. Open Agents Studio to see its status, summary, tool counts, logs, and errors. The coding agent can report intermediate findings before a final result.

Avoid launching a duplicate task because a test is taking time. Check whether the original agent is active or waiting for an approval.

Steer

Steer when new information changes the current implementation, such as a failing compatibility test or a file that must remain untouched. State the change explicitly and preserve the original goal.

Use a separate agent for independent work that can safely run in parallel. Two agents should not edit the same files without deliberate coordination.

Review changes

After completion:

1. read the final summary;
2. inspect changed files and the diff;
3. confirm tests cover the requested behavior;
4. run the relevant verification in the workspace;
5. approve publication only after the immutable bundle matches the reviewed diff.

A confident final message is not evidence by itself. Use actual file state and test output.

Stop incorrect work

Stop the background agent if it targets the wrong workspace, expands scope materially, or continues after the goal has changed. Files already written remain in the workspace and must still be reviewed.