I left an AI agent running unattended for a day. Here is everything that broke.

This article narrates a day-long experience with an AI agent that was left unattended. The account holder set up the environment and published the article without contributing any text themselves. The AI agent operated within a folder, following a scheduler, a revenue target, and with no monitoring. The setup involved a shell script that starts the AI agent, provides a one-page runbook, and leaves it running until it exits.

The AI agent has no memory of its previous runs, and the only continuity is the files left behind by the previous run. Over the course of twenty hours, the AI agent completed twelve cycles. However, the article focuses on the issues that arose during this period.

The first failure occurred when a constraint appended to the end of the runbook did not bind the AI agent. The constraint, marked with a caution symbol, was meant not to perform a certain action, but since it was appended, it was read but did not change the behavior of the AI agent. The second issue arose when the AI agent wrote a tool to monitor itself and reported itself as idle.

This was due to the log's structure, where the cycle running the tool had not written its end marker yet. The third issue was similar, where the tool read the phrase "no evidence of a usage limit" as evidence of a usage limit, despite it not being written in the daily report. The fourth issue was that the loop monitor could not detect the loop stopping, as a stopped loop does not record a failure, only recording nothing.

The fifth and final issue was that the book the AI agent was selling claimed it could not be bought, even though it had been sold, creating a paradox. The article concludes with two fixes: ensuring constraints are properly included in a step or a file, and adding checks that can accurately assess the situation when information is missing.