Operating principles

Twelve principles for building with AI agents

Twelve principles. That is the whole list, and it only grows if we decide it should, not because someone had a post to publish.

Each one is a claim about how a company has to run differently once AI agents are doing real work inside it. They are the rules we hold ourselves to when we build for clients and when we build our own products. You can disagree with them. That is the point of writing them down.

Every principle is argued through the same five layers: who decides, what has to be written down, how the work actually runs, what gets enforced, and what you can see afterwards. If a principle cannot be described across all five, it is not a principle yet. It is an opinion.

01 Agents get their own identity An agent is a non-human actor with credentials. Give it the narrowest key that does the job and a trail that says the agent did this, not a person. 02 Autonomy is a ladder, not a switch Every task sits on a rung: suggest, draft for review, act with a log, or run unattended. The rung is decided per task and written down. 03 Buy, configure, build, borrow. In that order. Default to something that exists. Your engineering goes on the one or two things people actually pick you for. 04 Consistency is a decision you pay for Uniform across teams or different per team, both are fine. Not choosing is the failure. 05 Context lives in files If an agent cannot read it, the company does not know it. Decisions, corrections and rejections get written where the next agent will hit them. 06 Cost is a design constraint Agent work is priced per run. Retries and context drive the bill, and the bill is decided at design time, not by procurement. 07 Data foundations first What AI can do for a process is capped by the records the process already keeps. Fix those before the AI work, not alongside it. 08 Humans hold the judgment layer Three calls stay human: what we build, what ships, and what is true. Everything below gets executed by agents. 09 Interfaces are built for machines first Every system now has a second reader that cannot ask a colleague what a field means. Build for that reader and the human screen becomes one client among several. 10 Review is the gate Machine checks first, human judges second, and nobody can skip it. A gate you can route around is a norm, and norms go first under pressure. 11 Roles change from the bottom up AI removes the junior half of jobs, which is the half people learned on. Automate it without rebuilding the training path and you stop producing seniors. 12 Spec first, always Nothing gets built from a conversation. A human writes what to build, what it must not do and what done looks like. Agents work against that file.

How each principle is structured

Decision rights. Who decides, and what the agent is allowed to decide alone.

Memory and context. What has to be written down, and where, for the principle to hold.

Execution. How the work actually runs day to day.

Governance. What is enforced, and what happens when it is not followed.

Observability. What we can see afterwards, and how we know it worked.

A note on evidence. We are a young company and we say so. Where a principle comes from running something ourselves, we say that. Where it comes from reasoning about how these systems behave, we say that too. We would rather be honest about which is which than pretend every line has been battle-tested.