Notes from the workspace.
Practical ideas for working with autonomous coding agents.
Give your agent a checkable outcome
A useful task has a clear finish line. Ask for a specific change, name the behavior you expect, and tell the agent how to verify it. A small task with a test is easier to assess than an open-ended request to improve an entire application.
Treat the command log as part of the deliverable
An explanation is useful; execution evidence is better. Read the actual command outputs and exit codes. Check that generated tests exercise the behavior that matters, and rerun important checks in your own environment before shipping.
Turn a good task into a reusable workflow
Once a prompt works well, save it as an automation. Keep the scope clear and use a workspace with the right repository. Start with manual runs before turning on a recurring schedule.