Agents keep getting better, but what they can accomplish depends on the arrangement around them. What a system can do lives in the relationships between its parts, not in any one part. The AI discourse is full of excellent parts; we are only beginning to build entire systems around them.

Over the last few months I turned my Obsidian vault into an external brain. An agent runs over a searchable corpus of years of ChatGPT conversations, my X posts and likes, ebook notes, and the conversations those notes spawn. The stack is opencode, DeepSeek v4.1 Flash (mostly), semantic search, and markdown. It has become a productive thinking environment, and a working example of a purposeful system of humans, machines, knowledge, tools, and feedback at the scale of one person.

This is a report from that small case, and a hypothesis that the same design logic holds when you add more people.

The agent

The agent is the common unit of design, and its vocabulary is frameworks, benchmarks, autonomy, memory, tool use. But an agent has a task, not a purpose. Purpose lives outside it, with the humans who decide which tasks are worth doing. You can build an excellent agent and still fail, because the failure is in the system.

In the systems I have built and supported, model capability is rarely the binding constraint. The harder problems are integration, authority, purpose, stale knowledge, and knowing whether the system is working. Those are organizational problems. Designing the agent in isolation optimizes a component and leaves the problems untouched.

The unit

AI-native design does not stop at the agent. It works on the whole system around it, where humans, machines, knowledge, tools, and feedback all shape what gets done and agency is spread across them. The unit is the human/machine information ecology; the parts are co-adapted. The first question is not “what should the agent do?” but “what purposeful system are we trying to create, and how should cognition, action, authority, memory, and judgment be shared among humans and machines?”

Purpose comes first, though purpose and architecture inform each other. Purpose is primary in authority. Someone has to own the direction, and that ownership cannot be delegated to the system being directed. Architectures shape what purposes you can see, and running systems teach you where stated purpose and behavior diverge. An unstated purpose still steers the system by whatever the parts reinforce, but no one is accountable for it.

An organization is a human/machine information ecology that persists toward a purpose. A company is one, but a single person augmented by models, notes, tools, and routines can exhibit the same properties. The smallest interesting organizational unit may be one human and one agent, and I suspect the design choices that work at that scale transfer to a larger group. The unit is the arrangement, not the participants.

Designing the ecology means asking a different set of questions: What knowledge can each participant access? What persists as memory? Which actions can models take directly, and which require authorization? Where does feedback about real outcomes come in? Where can participants change the arrangement when it is not producing what they want? Which decisions require human judgment? These are properties of the arrangement, not of the model.

The human

The question “what is a human?” keeps coming up, and it is a question about practice. What do humans actually do, and which parts of us are we handing over? We shape these systems and are shaped by them, in our habits, our attention, the questions we ask, and the ones we stop asking.

One part of that answer is goal formation, where explicitness is required for agents and humans to coordinate. Deciding what a system is for is where values, stakes, and context enter — the one part of the arrangement no model can supply. This is a normative claim: humans should set the direction. Models can contribute enormous cognition without supplying the source of organizational purpose. Venkatesh Rao makes a version of this point in Our Eukaryotic Moment, where humans supply stakes, desire, valuation, and contact with physical reality, which he calls liveness.

Variation is generated and selection is applied; culture is where that process becomes visible to itself. Models now generate variation at a scale previously unheard of, and humans do the selecting. Ask a model a bounded question and the answers converge; extend the inquiry and they diverge, because the differences come from our values and contexts. Models can increasingly help us select, but they cannot decide, on capability alone, what is worth selecting for.

The questions

  • How does migration differ from greenfield design? Most real organizations cannot start over. The interesting problem is redesigning a running system with agents inside it.
  • What does a company designed around this actually look like?
  • Where does the argument break? Holding at two scales is a hypothesis, not a given. The work is to find where the individual version fails to transfer to the organizational one, and vice versa.
  • What should be automated, and for whom? Some purposes are served by faster systems; others by systems their participants prefer, which may be slower and more expensive. That preference is a design input, not sentimentality. The pull to automate everything also runs against the fact that friction does productive work. In Accelerating Judgement I argued that production is accelerating faster than judgment, and that the organizational slack we remove was also involuntary deliberation. Bureaucracy is a rate limiter. Writing a proposal forces you to articulate the model. Explaining a system to another person exposes the gaps in your understanding. Some of that is waste. Some of it is the thinking happening while you do the work. When the answer is not obvious, the question loops back to purpose.

The claim

At the individual scale, I am already making these design choices, whether or not I call them organizational design. Every tool I adopt, every habit I build around it, every piece of knowledge I let go stale changes the ecology I think and work inside. The real choice is whether to do it on purpose.

Each of us holds the most context about our own goals and what matters to us. That context is the scarcest input to the whole process and the one no one else can supply on your behalf. The question I keep returning to — what is a human? — is answered in practice rather than in the abstract. It is answered by what each of us chooses to keep doing ourselves and what we hand to the machine.

Capability is becoming abundant. Judgment about what it is for is not.