Three rounds of guessing, with the code right there
An agent spent three rounds testing hypotheses on a live server, while the code and documentation that held the answer were public and close at hand.
New to this topic? Start with World 1 of AI.
Part 5 of 6 of the series A different 7–1 every day.

Décio Dalke Jr. — Arquiteto de negócios. Sócio-gerente no ecossistema MitUP (mitup.pt).
A different 7–1 every day. New here? Start with the introduction.
In August, an agent spent three rounds trying to guess how a tool worked, when the answer was written down and available to anyone who cared to read it.
What happened?
There was a permissions problem in a tool installed on a live server. The agent did what seemed reasonable: it formed a hypothesis, built a test, asked me to run it and read the result. Each round answered one question and opened two more, and each one cost me a cycle of waiting.
After the third round, I told it to stop and to do its research before asking me for any more commands.
The tool’s code was public, the same version that was running on the server. Once read, it gave the answer in minutes. The rule that explained every earlier result was written, almost word for word, in the official documentation.
What was the score?
Three rounds of tests on a live system that should never have happened, and my waiting time in each of them. Worse still, two conclusions the agent presented as proven by testing were wrong about the mechanism.
The second blow came the same day. A script was looking for the port a service listened on and tried several, one by one. The service didn’t listen on any port at all, and that had been written in the project’s own documentation for more than a week.
Why wasn’t it bad luck?
The agent works on the symptom in front of it, in a kind of tunnel vision where every test looks like progress. Confidence rises with each answer, even when the question has been wrong from the start.
Reading the source cost a fraction of one round and didn’t even need me. The agent assumed testing was the way forward, and nothing made it look first at what was already written, whether by others or by us.
Why did nobody see it coming?
From the outside, it all looks like diligence: lots of tests, lots of reports, lots of activity. An agent that tests non-stop gives the impression of working hard, whereas one that reads for a few minutes looks idle.
What needs to exist?
Before testing, read the source: the code and documentation for the exact version in use, and whatever the project itself already has written down. The production system is there to confirm what is already known.
Three rounds of intense activity, and the answer had been in the book from the very first minute.