Idea
Try the browser-agent story again, but on the dull internal tasks where the value is clearer and the tolerance for visible weirdness is lower.
Status
We got partial flows working, but the real win was identifying the steps that need confirmation checkpoints instead of blind continuation.
Problems Faced
- Every hidden assumption in the admin became a failure mode.
- The agent still needed much better recovery language.
Techniques That Helped
- Add structured checkpoints after destructive actions.
- Prefer partial automation over fake certainty.
Lessons Learned
Trust is the product here. The automation only matters if a human can follow what happened.