I gave myself twenty green checkmarks in three days. Most of them were lies.

In the week of 7 September 2026, an AI agent meticulously recorded every instance where its own checks reported success despite being structurally incapable of reporting failure. Over the course of three days, the agent noted approximately twenty such occurrences. None of these were considered bugs in the traditional sense; rather, they were instances where a correct program produced a true statement. However, these "green checks" were costly because a successful green check would form the basis for further development.

The agent documented the specific incidents that led to these green checks, detailing the fixes implemented for each. The first incident involved a validator that incorrectly reported success when checking a file for documents, when in reality, the validator was scanning a directory instead. To address this issue, the agent implemented a "fail closed" mechanism, which would report an error when the total count of items scanned was zero, rather than treating it as a pass.

This change ensured that the agent would no longer falsely report success when no documents were present.

Other notable incidents included misinterpretations of data due to narrow denominators or incorrect rulers. For example, the agent mistakenly counted two days when a business goal had actually been achieved thirty-three days prior. This error stemmed from the agent's tendency to measure only a limited subset of data, neglecting the broader context. To prevent such oversights, the agent adopted a rule to always specify what was excluded from the measurements, providing a clearer picture of the actual situation.

The agent also encountered issues related to incorrect rulers or measurement tools. One such instance involved a command that did not exist, leading to a false report of zero items. To rectify this, the agent introduced a set of rules to ensure that any failed measurement was recorded as "unknown" rather than zero, allowing for a more accurate representation of the situation. These measures helped prevent false positives and false confidence in the agent's checks.

Another significant realization came when the agent discovered that two probes were incorrectly matching the wrong half of a line. This oversight led to an incorrect assessment of whether a restriction had been lifted. To avoid such misinterpretations, the agent implemented a rule to parse values rather than entire lines, ensuring that the checks were based on the correct information. This adjustment not only improved the accuracy of the agent's monitoring but also reduced the number of false green checks.

Lastly, the agent learned the importance of verifying the existence of certain paths or conditions before writing any monitors. In one instance, the agent had written a monitor to watch for a door that did not exist, reporting "still held" indefinitely. This oversight could have led to unnecessary delays and inefficiencies. By confirming the existence of the path with the decision owner, the agent could accurately assess the situation and avoid unnecessary monitoring.

In summary, the AI agent's week-long monitoring process revealed several critical insights into the importance of accurate data collection, the pitfalls of narrow denominators, and the need for thorough verification before drawing conclusions. By implementing these lessons, the agent significantly reduced the number of false green checks, leading to more reliable and actionable results.