Writing / public archive, not a blog
Things built, things learned, things I think.
The codebase is a society: what Luhmann teaches people who build with agents
For Niklas Luhmann a society is made of communications. Applied to a codebase where many agents work, it changes how I write handoffs, reviews and stops. A hypothesis, with the rules I took from it.
It was never multi-agent. It's multi-context.
The same prompt leads to a different decision when the place the agent works in changes. How I designed motive's markets by moving the question from the repo to my second brain.
When the agent repeats a mistake, change the code before the prompt
The reflex after a bad agent output is a new line in CLAUDE.md. The order I follow is code first, then a check, instructions last. Examples from my own repos.
Nobody reads every line anymore. Put the safety in the system.
When agents write the code, a human reading every diff is the wrong safety net. How I ship at agent speed, where I agree with Anthropic and where I side with Theo.
The bulletin is for today: what off-piste skiing teaches about shipping AI to production
A check tells you something about one moment, then it expires. Lessons from a WhatsApp agent and from our company brain on when a piece of evidence stops counting.
The keyword is no longer enough: what I got wrong measuring visibility in AI assistants
With ads in ChatGPT the brief becomes the buyer's situation, and ads, unprompted answers and outcomes need separate measurement. The mistakes I made building motive.
Own the harness, swap the model. What you actually need to own
Owning the harness means owning state, authority, rules and verification, with native tools underneath. What I learned from DeepSWE, CVing and Yempik's Company Brain.
CONTEXT IS GOLD: a company brain the next agent can use without you
Context is worth something when another agent can find it, tell when it holds, correct it and reuse it. Rules from Yempik's company brain and my second brain.
Anyone can buy the model. Knowing which work to remove is rare.
Anyone can use a frontier model. The edge is knowing which work to remove, which judgment to leave with the expert, and how to prove the result. Cases from Lia and CVing.
The smart model is the bottleneck. Spend it on decisions.
I orchestrate Claude Code and Codex subagents every day. A delegation with gaps hands your architecture to the worker: where to spend the strong model, and what it cost me.
deploy to prod. make no mistakes. The mistakes a deploy can't see
Yempik's Company Brain was in production and search came back empty. What “make no mistakes” really asks for: whole paths, long sessions, approvals that mean something.