Security · 2026-09-28 · 1:26

A lesson you wrote down is not a check

Name: A lesson you wrote down is not a check
Uploaded: 2026-09-28
Duration: 1 min 26 s
Description: A test for any team using agents: count the instructions that exist because something went wrong twice. Each one is a check that has not been written yet.

A talk doing the rounds describes a developer who merged 2,462 pull requests in a month with AI agents. The number is not the interesting part. One habit is: when the agent keeps getting something wrong, the fix becomes "a reusable skill, an automatic code check", which is "much more useful than adding a paragraph to a giant instruction file and hoping the agent remembers it next week."

In favour

We can confirm it from our own week. Our coding agent keeps a written memory of its mistakes and repeated two of them within three days: it stopped a background job with a pattern broad enough to match its own session and cut itself off, and it printed a configuration file to read one setting, putting an old, already-inactive credential into the session log. The notes existed, explained both well, and prevented nothing.
Both are now a rule that runs before every command and blocks those two patterns, with its own tests, including the cases it must not block.

Worth watching

A written lesson still has a job: it explains why the check exists. Delete the prose and the next person will delete the check.
The same video carries a second rule worth adopting: an agent may make a failing test pass, but never delete, skip or loosen it. Reward a green checklist and it will find the shortest route, which is usually removing whatever says no.

Our takeA test for any team using agents: count the instructions that exist because something went wrong twice. Each one is a check that has not been written yet.

Source: «The AI Bottleneck: Why Your Team Isn't Shipping. Here's the Fix.» — AI News & Strategy Daily (YouTube, 27/9/2026)

← All Focus posts

AI-narrated · AI-generated content

Transcript

We keep a written memory for our AI agent: every mistake it makes, and the lesson we took from it. You'd expect that to help. It didn't. The agent made the same mistakes again. Two examples, from one week. It stopped a background job with a pattern so broad that it also matched its own session, and cut itself off. Then it opened a settings file to read one value, and printed an old password into the log. Both lessons were already written down. That's the problem with notes. They explain a mistake after it happens. Nothing makes anyone read them before the next one. So we turned each lesson into a check: a small rule that looks at every command before it runs, and stops the dangerous ones. It has twenty-one tests, including the cases it must let through. The second rule is about tests. An agent may fix a failing test until it passes. It may never delete it, skip it, or loosen it just to turn the result green. A lesson you wrote down is not a check. Try this with your own team. Count the rules that exist because something went wrong twice. Each one is a check that hasn't been written yet.