This week I asked one of our agents to fix a cluttered page. I named two sections. It fixed those two and left the same problem in the section beside them.
So I asked for the guidance to be improved, and the agent added rules in five places. One of them pushed our company philosophy, the document every agent starts from, over the limit for loading at the start of a session. It stayed over for three minutes, until the next fix. An instruction meant to keep agents on the mission had crowded the mission out. I stopped asking for rules and asked for an evaluation instead.
What the evaluation showed
Hyperdrift runs on agents. They build, release, write and follow up across our apps. Their guidance had grown one correction at a time, and when we counted, it held 324 separate “do not” instructions across 67 files.
A rule tells an agent what to do. An eval tells me what it did. So before writing anything new, we checked whether the guidance we already had was opened. We read fourteen days of sessions from one machine, with both tools working in the same repository.
In open conversations, Claude opened one of our written methods in 167 of 307 sessions, and Codex in 39 of 41. Codex’s 30 scheduled routines and all 77 of its helper sessions opened one too. We did not count Claude’s helpers.
The work itself told the same story. When a Claude conversation changed a page’s interface, it had opened a design method in 23 of 48 sessions. Before a release, it had opened the release method in 27 of 78.
Opening a method is not the same as following it, the two samples are different sizes, and I can’t yet say why the tools differ. What the counts do tell me is that nearly half of one tool’s conversations opened none of the methods we had written. No rule fixes that, because a rule is one more thing to open.
There was more. Two of the three places our agents run held a copy of the philosophy eleven days old, without the paragraphs added the night before. For one of the two, that is what its release record shows; we could not read the file itself. And the weekly review meant to catch problems like these had last run on 2 November. It had refused to start on its last two runs.
Why we deleted rules
Those 324 rules had accumulated as replies to mistakes. Together they did to the guidance what the agent had done to my page: they buried the point. What I told the agent was: “It should just do the right things if it follows the philosophy.” And: “I’d rather flag failures to sharpen alignment than add rules everywhere.”
We kept a rule when it named a specific fact or a decision I had made. We removed it when it described character, because character belongs in the mission. Thirty-seven prohibitions went first, then five whole files and about fifty more lines.
My bet is that an agent that knows the mission needs fewer rules, and makes a better call when the rules are silent. I haven’t measured that yet. That is the job of the evals: the same counts now go into our weekly review, and each failure it flags is a reason to sharpen the mission before anyone writes a rule. Part three covers how we run them.
The same lesson in a human team
In part one I wrote that poor communication is the first thing that takes a team away from its founder’s vision: people work on initiatives that each make sense in isolation. Our agents showed me the same pattern, with counts attached. In a company it would be a handbook half the team never opens, a plan that changed while some people work from the old copy, and a review that stops happening.
I don’t think a longer handbook fixes that. What works is people who know what the company is for, who heard it recently, and whose work someone connects back to it. I believe that is also where motivation comes from. Someone working in isolation from a list does what the ticket says. Someone who can see how today’s work serves the vision brings judgement to it, and keeps going.
The idea Traction started from
Traction started from that idea, for founders. A motivated founder building alone has the vision, but not the team that keeps each day’s work connected to it. In Traction, an AI Growth Manager runs the work while you lead. It reads your website, says what works, what’s in the way and the move to make first, and nothing goes out without your yes.
If you run agents or a team, try one thing this week: before you add a rule, run the smallest eval you can. Count how often the last rules were opened.
