Five manual checks before trusting an agent actually finished its job
Kodenstudio · reddit · 2026-11-03
The author shares five manual checks for silent failures in long-running agent/cron systems, after a missing WHERE clause cost four days: verify lease claims (two jobs died before claiming one and sat silently for four days), delete a line or flip a boolean to prove tests aren't decoration, read the lastsuccess timestamp instead of the status column (showed 'running' but last write was 96h ago), open logs of anything that stopped complaining (a refactored-away parser went unnoticed for three weeks), and re-run yesterday's successful job to catch unsafe retries that write duplicate rows. Final test: if the answer to 'how long until we notice?' is 'when someone complains,' nothing is watching. Five minutes a week.
More from coding & agent
- Tip: Have your agent export GitHub Actions CI timelines as Perfetto .pftrace files — vincent_koc · 2026-11-03
- Human mops while their paid AI agent renders an After Effects project solo — ZeroStateReflex · 2026-11-03
- 9 GitHub skills repos that turn Claude into a real work assistant — nikola_mr64990 · 2026-11-03
- Claude Code caches are model-specific: switching models, compacting and idle time all bust the cache — shoumikgoswami · 2026-11-03
- I made my agent ask teammates' agents directly instead of waiting on me — Upset_Ad_2985 · 2026-11-03
- LangChain founder asks: is agent-level memory just 'skills' with a trace-updating process? — hwchase17 · 2026-11-03