Walkthrough

Stopping, Steering, and Retrying

Three levers for a run going wrong: comment to steer, stop outright, or retry with a sharper brief.

Long agent runs go wrong in two ways: drift (working hard on the wrong thing) and stall (stuck on an environment problem, or looping on the same failing step). The Manager gives you three levers. Steer with a comment when the correction fits in a sentence. Stop when the premise was wrong. Retry with a rewritten brief when the failure taught you what the brief was missing.

Steps · 0 / 3 done
  1. Steer mid-run with a comment

    Drift you can correct in one sentence doesn't deserve a restart — comment on the task list or plan and let the agent adjust with its context intact.

    Comment on the plan: "Stop adding the settings page — out of scope. Finish the limiter, flag the settings idea as a follow-up."
    VerifyThe agent acknowledges the comment, updates its task list, and the diff stops growing in the wrong direction.
  2. Stop cleanly and inspect

    The stop control halts the agent without vaporizing its work — edits stay on the branch/worktree, so you can read the partial diff before deciding what's salvageable.

    git status && git diff --stat
    VerifyYou can see exactly what the stopped agent changed, keep what's good, and discard the rest.
  3. Make failure cheap with git

    Because agents work on branches, a bad run costs one command to erase. This is why per-agent branches are non-negotiable — cheap failure is what makes bold delegation rational.

    git restore . && git clean -fd   # or just delete the agent's branch
    VerifyClean tree. Rewrite the brief with what the failure taught you, and relaunch.
Check your understanding
Q1. The agent is building the wrong extra feature, but the core work is on track. Steer, stop, or retry?
· Tick off the 3 step(s) above.
· Score 100% on the quiz.