Agents feel less reliable since September model updates, devs log failure rates to find out

john_snow_21 · reddit · 2026-09-21

A developer reports that after the recent wave of model updates, previously rock-solid agent workflows started missing steps, with more retries and small errors slowly eroding trust.

They can't tell whether it's real degradation, prompt drift, or confirmation bias amplified by widespread claims that models quietly got worse. They've begun logging per-flow failure rates so the next update cycle brings data instead of a feeling. The thread asks whether others are genuinely seeing more agent flakiness or just internet panic.

Related event: Developers report agents getting flaky after September model updates(2 posts)→

Original post →

More from coding & agent

coding & agent channel →