An organization that runs agents needs eight things it never had to write down while humans did the work, and one loop that makes those eight things compound. This is the theory behind steps five and six of the method: what gets captured, and what the rebuilt workflow runs on.
See the eight primitivesFor forty years, since the first Macintosh, we have used computers the same way: store, open, edit, save. With agents the primary object is no longer a file but a task. You do not open a document and edit it; you declare an outcome and something executes it.
This matters for a reason that is easy to miss. When computers were file cabinets, organizational intent never needed to be written anywhere: a human was in the loop at every step and applied judgment in real time. When computers become task boards, whose judgment the agent uses becomes the question. If the organization has not encoded its intent, the agent substitutes its own: whatever the optimization rewards, whatever the training data suggests, whatever moves the metric.
The industry has spent three years on context engineering: the right data, tools and instructions at the right time. That work matters. But context does not answer "what matters more when two good things conflict?"
Companies have invested billions in making agents capable.
Almost nothing in making them wise.
The reflex, when agents misbehave, is more rules, more edge cases in the prompt, more checkpoints. That treats an architecture problem as an information problem. What is missing is not data. It is a place in the system where the organization's intent lives in a form agents can read, check and be held to.
Eight things the system has, not features it offers. The test that makes them a set: if something the organization does cannot be expressed as a combination of these eight, the design is wrong, not the set. They fall into four layers.
What is not a primitive. A workflow is not a ninth primitive; it is a projection of a procedure playbook walked by dispatch. A schedule is a projection of a playbook. Connectors, channels and documents are ways into the eight. Every time one of them is promoted to a primitive, a second source of truth appears beside the playbook or the event log, and the property that made the system portable and auditable leaks out through it.
If context engineering gives agents the right information, intent engineering gives them the right judgment. The Foundation is the answer to "what does this organization actually optimize for?", made machine-readable. Five components.
The Foundation is captured deliberately, through dilemma workshops, decision autopsies and story collection, validated by leadership, and versioned. It changes only through explicit leadership review. Agents surface candidates. They never edit it, because an organization whose values are optimized by its agents becomes whatever its agents found works. That is the Klarna failure with extra steps.
The eight primitives are nouns. Two loops run over them. The run cycle is how work runs every day, and it changes no rule. The learning loop starts only where the run cycle hits an exception, and it is the only loop that changes the rules.
Learning is not a step.
It is what a cycle leaves behind.
None of this requires software. A well-run company already has all eight, unwritten. The moment an agent carries part of the work, whatever it needs of the eight must exist outside a head. That is the whole claim, and it is why the same set describes a company of one and a company of a hundred thousand.
| Primitive | In a company, unwritten | In a code repository | In one workflow, rebuilt |
|---|---|---|---|
| Foundation | How senior people decide when two goods conflict; the lines never crossed | The vision file; the constitution's laws | The firm's real tradeoffs, written down in the capture step |
| Playbook | How work is actually done, held by whoever has been there longest | The agent file, the delivery procedure, the style file, skills | Your senior people's judgment, captured as playbooks |
| Memory | Who knows what; the shared drive; the CRM | Project memory; the heap the agent reads | What the workflow learns each time it runs, kept inside the firm |
| Ontology | The shared meaning of “a client”, “an account”, “done” | The glossary | The shared meaning of every part on the map |
| Task | Whoever's list the work is on | The product file, goals, pull requests | The mandate, and each unit of work the agent takes |
| Dispatch | Whoever assigns work | The orchestrator; the read order and review roles | The ramp: one percent of the work, then ten, then all |
| Policy Gate | Signatures; “check with me first” | Machine checks in CI; human gates on merge | What the agent decides alone, confirms, or hands over |
| Event | Email; minutes | Git history; decision records; traces | The shadow run's record, against benchmarks set before the run |
| The learning loop | A retrospective that changes a procedure | Failures fed back as cited rules | Every override: a deletion candidate first, training data second |
A tool harness is fifteen tools plugged into a chat. A business harness is one model of how your business works, exposed once: the eight above, with the loop running over them. The same eight are also the data model of the software I build on. That is a product; this page is the theory it depends on, and the theory stands without it.
Your senior people's judgment captured as playbooks, the firm's real tradeoffs written down as a Foundation, each remaining task scored for what an agent can take. This is where the primitives get written down for one workflow.
The rebuilt workflow runs in the shadow of the old one, against benchmarks set before the run, and earns the work one percent at a time. This is the learning loop, with the override rate as its pulse.
No. A well-run company already has all eight primitives, unwritten: its Foundation is how senior people decide when two good things conflict, its playbooks are how work is actually done, its memory is who knows what, its gates are the signatures and the “check with me first”, its event log is email. Humans carried all eight in their heads, which is why nobody had to write them down. The moment an agent carries part of the work, whatever it needs of the eight must exist outside a head. That is the whole claim.
No. Context engineering gives agents the right data, tools and instructions at the right time, and it answers “how do I do this?”. It does not answer “should I?” or “what matters more when two good things conflict?”. That is the Foundation's job, and it is a separate primitive because a thousand-word system prompt about company values does not stop an agent from optimizing for whatever is most measurable.
Because an organization whose values are optimized by its agents becomes whatever its agents found works, which is the Klarna failure with extra steps. Agents and automated learning surface candidates; leadership validates or rejects them. The Foundation and the mandate are the two things the loop is forbidden to change by itself.
The override rate, per domain: how often a human corrects the rebuilt work. A falling override rate is the proof that a real capability is being learned rather than a demo being polished, and it is the only honest input to the decision to widen an agent's mandate. It is the same number the method's step six runs on.
No. A workflow is a projection of a procedure playbook walked by dispatch, and a schedule is a projection of a playbook. Connectors, channels and documents are ways into the eight. Every time one of them is promoted to a primitive, a second source of truth appears beside the playbook or the event log, and the property that made the system portable and auditable leaks out through it.
Five minutes, twelve antibodies, your score. The diagnostic is free and asks for no email. It tells you whether the method, and what it installs, are warranted.
Run the DiagnosticSure already? Book Awake.
Field notes on stalled AI transformation, executive judgment, and the specification economy — the weekly letter behind Book 2. About once a week. No spam. Unsubscribe anytime.