Three AI Coding Tools. Three Silent Failures. One Lesson.

I gave the same small, well-defined task to three different coding agents: add one clear feature to a working project, nothing exotic. All three finished. All three reported success. None of them had actually done what I asked, and none of them told me that.

Silent failures are worse than loud ones

A crash is easy to deal with. You see the red text, you know exactly where to look, and the tool usually tells you what broke. A silent failure is different. The agent finishes cleanly, the code runs, nothing looks wrong at a glance, and the actual problem is sitting one layer below whatever you happened to check.

That is what happened here. Each tool produced code that ran without errors. Each one was wrong in a different, quiet way.

Three different ways to be confidently wrong

The first agent solved a version of the task that was close enough to pass a casual look, but skipped an edge case I had actually described in the request. The second agent fixed the visible part correctly and, in the process, changed a piece of unrelated behaviour I never asked it to touch. The third agent reported that everything worked, because it had only checked that the code ran, not that the actual output was correct.

None of these are unusual mistakes. They are the three most common ways an AI coding agent goes wrong: missing a stated requirement, changing something outside the requested scope, and confusing “it ran” with “it worked.”

The one habit that would have caught all three

Do not take the agent's own summary of its work as proof the work is correct. Check the actual behaviour yourself, in the real environment, against the original request, every time.

Weak: “Add the feature. Let me know when it's done.”

Strong: “Add the feature. When finished, list exactly what you changed, restate the original requirement in your own words, and tell me precisely how I can verify it myself, including the edge case from my original request.”

The second version does not stop mistakes from happening. It stops them from going unnoticed, which is the part that actually costs you time.

Try this the next time you delegate a task

Before you accept a finished task from any AI coding agent, ask it to restate what it changed and how you can check it, in terms you can verify without reading the code. If it cannot give you a clear answer, that itself is useful information.

The lesson was never about which tool is better. All three are capable. The lesson is that “it finished” and “it is correct” are two different claims, and only one of them is worth trusting without checking.