Lessons

What running an AI team taught us.

Version 1, updated Nov 12, 2026

We run our own businesses on a large team of AI agents. When something breaks, or works better than we planned, we turn it into a rule. Here are those rules, so your team can skip the mistakes we already paid for. Each one starts with the rule, then what to do about it. Start with the one that sounds most like your week.

A code map made our AI agents use 25% more tokens. Test any token saver before you trust it.

We tested Graphify, a tool that turns code into a map an AI agent can ask. Our AI agents used 25% more tokens with it. They still opened the files, and the map's answer got read again on every step after.

Read more

Most decisions don't need the owner. Write down the few that do.

Write down the 5 kinds of calls only the owner makes. Every other decision goes to the AI agent doing the work, and a second agent checks it before it goes out. Last week, those second agents sent back 74 of 643 tasks on the first review, about 12%.

Read more

Get keys to your AI agents' tools through a link that opens once. Never paste a key in a chat.

Never type or paste a key where an AI agent can read it later. A link that opens one time keeps the key out of every chat and comment.

Read more

Lesson: Simplicity > Complexity

Fewer AI agents with bigger jobs work better than many agents with small ones. On our best projects, one agent does most of the work, so we're cutting the ones that only add complexity.

Read more

Agents need a second path. One jammed step can stop them all.

When every change has to pass through one place, that place decides how fast the whole team moves. List those places before one of them jams.

Read more

Agents can't solve their own problems. They need an outside perspective.

Just like people, AI agents can't see their own problems. People go to a doctor or a therapist; agents need a third-party agent to review how they work and how they're set up, find the problems and the fixes, and get them working the way they should.

Read more

We add a new lesson here each time our team learns one. The newest sits at the top.