My Routine Said It Ran. It Was Lying.

I run an AI system. It follows a schedule to maintain itself. One routine runs twice a week to save a file.

The scheduler said it ran. The timestamp was there. The status was green.

The file did not exist. It was not in the wrong folder. It was nowhere.

People miss this when they build autonomous agents. "It ran" and "it worked" are not the same. Most dashboards only check if it ran.

A scheduler tells you a process started. It tells you nothing about the result. My routine started, hit an error reading a missing file, and stopped. It did not crash. There was no red error message. It ran, but it produced nothing.

I stopped looking at the status. I went to the disk. I used three checks:

  • Does the output file exist right now?
  • Is the file fresh and not empty?
  • If the file is missing, read the raw run log. Do not read the summary. Read the actual transcript of every tool call.

The summary said the routine was "episodic." The transcript told the truth. The agent tried to read a memory file. The file did not exist. The agent never reached the step to create the file. It made zero write calls.

The summary would have led me to the wrong fix.

If you run autonomous systems, remember these rules:

  • "Ran" is not "worked." Health is the file itself. The file must exist, be fresh, and not be empty.
  • The spec is a hypothesis. The disk is the fact. When they disagree, trust the disk.
  • Read the raw log. The summary is written by the same system that failed. The transcript shows which tools actually ran.

The fix was simple. I made the routine write its file first. It now writes the file before any other step.

Most agent problems are not reasoning failures. They are plumbing failures.

Source: https://dev.to/eltony_lfgi/my-routine-said-it-ran-it-was-lying-2gfb

Optional learning community: https://t.me/GyaanSetuAi