Agent fixed what its reviewer flagged—then the next review passed the new bug it created

Fit-Ant2238 · reddit · 2026-09-27

The author ran a fully autonomous agent workflow building pitch slides for a fictional board game, using a skill pack (SenseNova-Skills) that screenshots every slide and sends them to reviewer sub-agents.

The first pass worked: sub-agents caught a route line crossing the cover title and beacon towers floating above hills, and the main agent fixed both. But when told the two towers on the last slide were too small, the agent enlarged them straight into the slide number and footer text.

The next review round—where a fresh sub-agent got the list of fixes and was asked to confirm each one—declared the towers plainly visible and reported no overlapping text; the automated render check passed too. The shipped deck has the slide number printed across one tower.

The author's takeaway: a reviewer that knows what got fixed tends to check the fix and little around it. He asks whether people give review agents the change list or make them review cold.

Original post →

More from coding & agent

coding & agent channel →