Agent got rescued by hand? Route that fix back into skill versioning

Jimcy-Maffesoli · reddit · 2026-09-14

A look at where manual corrections to stuck agents should go: if the fix stays in someone's head, the next run repeats the same blind spot. Reef's harness tutorial offers a feedback loop — record the request, attach a score and correction to the run's receipt, let the model propose a skill change, then A/B test current vs. changed harness on the same tasks before publishing a version. Humans still define what a good fix means and install the chosen version, but interventions become input to the next iteration instead of vanishing when the ticket closes.

Original post →

More from coding & agent

coding & agent channel →