AgentSeism: open-source statistical CI for deciding when an agent truly regressed

puppy_lover_2021 · reddit · 2026-09-27

A developer building CI for agents shares hard-won lessons: an unchanged agent fluctuated from 92% to 88% success (a false alarm if blocked), while a deliberately degraded agent dropped 92% to 52% yet passed a naive statistical rule because damage concentrated in a few tasks. He argues agent PRs should be treated as experiments, not deterministic tests, separating broad reliability drift from single-capability collapse. The open-sourced AgentSeism uses repeated runs, practical effect thresholds, uncertainty, multiplicity across capabilities, and an explicit "insufficient evidence" state. Blog and repo included, seeking feedback on merge-blocking criteria.

Original post →

More from coding & agent

coding & agent channel →