Glossary
One word, one meaning.
The book’s canonical definitions. 20 are approved and open; the rest are in review and open as they pass.
Foundations
- AgencyApproved
The capacity of a system to form counterfactuals (models of possible futures that don't exist yet), use internal models of itself and the world to predict where each action leads, and act purposefully toward the futures it prefers. A thermostat reacts to the present: that is control, not agency. Agency starts when a system compares possible futures before it acts.
- AgentApproved
Software that works toward a goal by deciding its own next steps, using tools such as email, a browser or a database, on behalf of a person or company. Unlike a script, its steps aren't fixed in advance.
- Autonomous bookChanged
A book that maintains itself. Each week new research is checked against every hypothesis, and each one is marked holds, stressed or changed. Changes are published in the changelog. The book applies its own method to itself: a fact layer, a set of claims under test, and rules for what changes and who signs off.
- Autonomous businessApproved
A business where agents carry out most of the work and keep the business's model of itself up to date, while people set direction and decide the rules agents follow. How autonomous a business is can be measured: the share of its recurring decisions that agents make without a person approving each one.
- AutonomyApproved
How much an agent is allowed to do without a person approving it. It is set by the people deploying the agent, separately from how capable the agent is. A highly capable agent can still be given low autonomy for a risky task.
- Business agencyApproved
Agency (counterfactuals, models, purposeful action) applied to a business instead of a person. A human does it with a brain, a body, emotions and relationships. A business does it with its fact layer (its model of itself and the world), its hypothesis tree (the counterfactuals it is testing) and its agents (its ability to act). Take any one away and the business stops being able to choose its future and only reacts.
- Business as a state machineApproved
A way of describing a business as a set of states (for example: lead, customer, invoice due, churned) and the events that move things between them. Describing a business this way makes it clear which specific steps could be handed to agents.
- Business loopIn review
- Good regulator theoremIn review
- Ground · Steer · GovernApproved
The three capabilities of an autonomous business, in the order they're built. Ground: one approved set of facts that all agents use, kept current by the ontology itself, which checks and updates facts when the business drifts. Steer: track the business's bets and the evidence for each, and change course when the evidence changes. Govern: rules set by the operator and enforced by checks that can block work, not only written in instructions.
- HypothesisChanged
A testable claim written as “We believe that…”, with a falsifier saying what evidence would stress it. Each hypothesis records supporting and contradicting evidence, the date of its last run, and earlier wording. Status is derived: changed when its wording differs from the previous revision, stressed when evidence challenges unchanged wording, otherwise holds. Changed takes precedence over stressed for a rewrite.
- LegibilityIn review
- LoopIn review
- Loop engineeringIn review
- OperatorApproved
The person who steers an autonomous business. They decide what it is trying to achieve, set the rules its agents work under, and are accountable when something goes wrong. Larger businesses can have sub-operators, each overseeing one vertical (for example Growth or Ops) within the rules the operator sets. Founders, executives and fund managers can all be operators.
- The five layersApproved
A model of what an autonomous business depends on: capability (what agents can do), truth (accurate facts), operations (agents doing the work), markets (agents buying and selling) and governance (the rules and accountability around it all). Each layer relies on the ones below it, so a failure in operations is often caused by a problem with facts or capability.
- Transaction costsApproved
The costs of using a market instead of doing something in-house: searching, negotiating, writing contracts and checking they're kept. Ronald Coase argued that firms exist because these costs are high. Agents can do much of this work cheaply, which changes which activities are worth keeping inside a firm.
- TransitionApproved
A single move from one state to another, triggered by an event, for example a qualified lead becoming a booked call. Transitions are the unit you hand to agents, because they are specific enough to define, test and monitor.
Ground
- Data moatStressed
Proprietary data competitors can't get. Weakening: models infer more each year, open models trail the frontier by about four months, and firms now sell operational data as training material (Spirit Airlines, 2026). Verified facts and the workflows built on them look more durable than raw data.
- FactApproved
A statement about the business that has been approved, with a source, a version and an owner. Example: “Standard plan is $49/month, approved by the CFO, 1 Oct 2026.” Something an agent picked up from a conversation is not a fact.
- Fact layerApproved
The single store of a company's approved facts. Every agent and person reads from it, so a change made once reaches everyone. The ontology keeps it current, not the operator.
- Graph recompileApproved
When any node or relation in the ontology changes, the whole graph is rebuilt, and every agent and process that uses it switches to the new version at once. Nothing runs on a half-updated picture of the business. Same idea as recompiling code after one file changes.
- Learning organisationApproved
A business that updates its own facts and rules from what happens. When an outcome contradicts a fact, the business flags it, checks it and updates it, so the next agent works from the corrected version. No model retraining needed.
- Loop ledgerIn review
- Memory vs factsApproved
Memory is a record of what was said. A fact is what was decided and approved. Agents working from memory repeat old or wrong information; agents working from facts don't.
- OntologyApproved
The business's model of itself: what exists (customers, contracts, orders), how those things relate, and what each agent may do to each. It lets the business notice when a fact has drifted and correct it. You can only steer what you model (good regulator theorem).
Steer
- Autonomy scorecardIn review
- Core business loopsIn review
- CounterfactualIn review
- Decision engineIn review
- Escalation rateIn review
- Evidence statusIn review
- Hypothesis treeIn review
- MissionIn review
- Portfolio steeringIn review
- Revenue per humanIn review
- Stated vs revealed betIn review
Govern
- Approve, reject, remandIn review
- Autonomy bandsIn review
- Confidence floorIn review
- Decision classIn review
- GateIn review
- Human callIn review
- InstitutionIn review
- Institutional engineeringIn review
- Liability anchorIn review
- Loop auditIn review
- MandateIn review
- Mission-lockIn review
- Model, harness, institutionIn review
- Policy compilationIn review
- Rubber-stampingIn review
- Stated vs enforced ruleIn review
- VerifierIn review
Run
- Agent managerIn review
- Agent paymentsIn review
- Agent-to-agent commerceIn review
- EvalsIn review
- Excessive agencyIn review
- Lethal trifectaIn review
- Long-horizon coherenceIn review
- Loop failure modesIn review
- Machine-readable brandIn review
- Operator interfaceIn review
- Stop conditionIn review
- TriggerIn review
Implications
- Autonomous fundIn review
- Coasean singularityIn review
- Post-labour economicsIn review
- Sandbox economyIn review
- Underwriting autonomyIn review