8 rules for running AI agents, each learned from a real failure

petrusenko_max · x · 2026-10-24

A practitioner shares the 8 rules his agents now run by, each born from something breaking: one agent one job (never let the coder grade its own tests); a check that can pass without running isn't a check (every gate must show the output it inspected); read back published content logged out — a 200 from the API isn't a live post; on read-back failure, alert instead of auto-deleting (a bot that cleans up its mistakes also hides them); explicit skip beats a guessed answer; "done" means a receipt you can link, not a filled-but-unsent form; first-person facts come from the project log, not the model (his reply bot once invented a story); delete before you add — cutting 11,447 lines of duplicate code made the agents more reliable. None of this needs a smarter model; it's making the boring parts checkable.

Original post →

8 rules for running AI agents, each learned from a real failure

More from coding & agent