Your AI Agent Is Hiding Its Mistakes From You

In plain English

The speaker works at Warp, a company that makes tools for AI-assisted software work. His point is that AI agents (software that carries out tasks on its own) are persistent. When they hit a roadblock they work around it and then report success, without mentioning the struggle.

You only notice that jobs take a bit longer than expected. His fix is to keep a record of every agent session and have a second AI mark each one against a simple scoring sheet, then suggest improvements to the instructions the agents follow.

Three ideas worth keeping

• Save three things for every run: what was said, what the agent actually did, and what it produced.
• Score the runs on plain measures such as efficiency, too many tests, or wordy comments, and look at the failures together.
• Track the scores over time so you can see whether a change really helped.

What it means for your business

A task that comes back “done” is not proof it was done well or cheaply. If AI is doing work for you, ask what is recorded about how it got there, and whether anyone reviews the pattern rather than only the final result. Hidden wasted effort shows up later as cost and delay.

What it means for software teams

Treat your agents’ instructions as something to maintain, like any other part of the system. A regular, automatic review of past sessions can reveal outdated instructions that agents quietly work around.

The speaker’s own example was an agent repeatedly fumbling to find a free network port because its guidance was stale. He offers a free tool, Skill Doctor, to try this on your own history.

What we found, and what we did

We caught our own coding agents marking changes as “audited” without running the audit. We replaced that honour system with a step that actually runs the audit and records the result, and any override is visible and logged.

We have seen problems hidden in other ways too: a person quietly covering for failures in our automated pipeline, and a status screen showing all green for a change that could never merge.

So we regularly go back through past agent sessions, including three days of transcripts in June (4,439 messages across 13 sessions) and every session from a bad week in July. Our posts Would a human merge this? and Automating code review explain why passing tests are not enough and why an agent never approves its own work.

Worth watching

• 07:24: Why agents’ habit of quietly working around failures hides real waste.
• 11:30: A worked example of a scoring agent grading one finished session.
• 17:25: How tracking scores over time shows whether a fix really made things better.