Changelog

What changed, and why.

The public record of the book. Every hypothesis revision keeps its wording, evidence and date.

Holds no challenging evidence in the latest runStressed evidence challenges the current wordingChanged wording changed; previous wording is preserved

Versioning

v1.1 rewrites the launch edition’s 21 claims as hypothesis drafts with explicit falsifiers. This is an editorial revision, not a new research run. The v1.0 wording and evidence remain below. A rewrite is Changed even when contradicting evidence remains; the next run with unchanged wording is Stressed if that evidence still challenges it. Hypothesis revisions increment the minor version; restructuring a part increments the major version.

64 entries

  1. v1.2New
    Why a living textbook

    Added an in-chapter UFAST application with documented tool setup, an agent interview command and a completion test. Existing hypothesis evidence is unchanged.

  2. v1.2New
    The business as a state machine

    Added an in-chapter UFAST application with documented tool setup, an agent interview command and a completion test. Existing hypothesis evidence is unchanged.

  3. v1.2New
    The fact layer

    Added an in-chapter UFAST application with documented tool setup, an agent interview command and a completion test. Existing hypothesis evidence is unchanged.

  4. v1.2New
    Learning organisation

    Added an in-chapter UFAST application with documented tool setup, an agent interview command and a completion test. Existing hypothesis evidence is unchanged.

  5. v1.2New
    Data as moat

    Published chapter prose with researched UFAST application, tool setup, interview command and completion test.

  6. v1.2New
    Discovery

    Published chapter prose with researched UFAST application, tool setup, interview command and completion test.

  7. v1.2New
    Measuring autonomy

    Published chapter prose with researched UFAST application, tool setup, interview command and completion test.

  8. v1.2New
    Redefining the operator

    Published chapter prose with researched UFAST application, tool setup, interview command and completion test.

  9. v1.2New
    Trust and brand

    Published chapter prose with researched UFAST application, tool setup, interview command and completion test.

  10. v1.2New
    Digital businesses

    Published chapter prose with researched UFAST application, tool setup, interview command and completion test.

  11. v1.2New
    Skills and playbooks

    Added to Run with researched UFAST application and agent interview command.

  12. v1.2New
    Security and local models

    Published chapter prose with researched UFAST application, tool setup, interview command and completion test.

  13. v1.2New
    Agent evaluation and reliability

    Published chapter prose with researched UFAST application, tool setup, interview command and completion test.

  14. v1.2New
    Growth loops

    Added to Run with researched UFAST application and agent interview command.

  15. CP-2Changed
    Glossary · Fact layer

    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.

    Previous wording: 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; that is not the operator's job.

    Fact layer: tighten the last sentence (applier test). End-to-end test of the applier: a small wording change to a published term, with the same meaning.

  16. v1.1Changed
    Why a living textbook

    We believe that a static book about agents goes out of date within one capability doubling.

    Previous wording: A static book about agents goes out of date within one capability doubling.

    Stressed if a capability doubling leaves the book’s practical recommendations accurate without revision.

    Editorial hypothesis draft; existing evidence carried forward, not a new research run.

    Supporting evidence: METR — Time Horizons

    Contradicting evidence: None recorded.

  17. v1.1Changed
    Why a living textbook

    We believe that research updates can be automated, while publication decisions require human review.

    Previous wording: Updates can be automated. The decision to publish them cannot.

    Stressed if automated publication matches human review for accuracy, provenance and correction rates on comparable research updates.

    The counterpoint source supports this hypothesis: quality assurance is the part automation has not solved. Editorial hypothesis draft; existing evidence carried forward, not a new research run.

    Supporting evidence: ISPOR 2026 — GenAI framework for living reviews; Song et al. — Phases of Living Evidence Synthesis Using AI (JMIR); ISPOR 2026 — Living literature reviews: maintaining quality

    Contradicting evidence: None recorded.

  18. v1.1Changed
    Why a living textbook

    We believe that publishing what stopped being true helps readers update their decisions.

    Previous wording: Publish what stopped being true, not only what is new.

    Stressed if readers given a revision history make no better-informed decisions than readers shown only the latest text.

    Editorial hypothesis draft; existing evidence carried forward, not a new research run.

    Supporting evidence: Elliott et al. — Living systematic reviews; CASRAI — Definition: living review

    Contradicting evidence: None recorded.

  19. v1.1Changed
    History of the firm

    We believe that firms exist because using the market has transaction costs.

    Previous wording: Firms exist because using the market costs something.

    Stressed if firm boundaries do not respond to changes in search, contracting or enforcement costs, after other factors are accounted for.

    Editorial hypothesis draft; existing evidence carried forward, not a new research run.

    Supporting evidence: Coase — The Nature of the Firm; Williamson — Markets and Hierarchies

    Contradicting evidence: None recorded.

  20. v1.1Changed
    History of the firm

    We believe that agents lower search, negotiation and enforcement costs enough to move more work outside the firm.

    Previous wording: Agents cut the cost of searching, negotiating and enforcing, so more work can move outside the firm.

    Stressed if platform fees, dependence or verification costs outweigh the coordination savings from agents.

    Platform dependence can raise the external costs that agents lower inside the firm. Editorial hypothesis draft; existing evidence carried forward, not a new research run.

    Supporting evidence: Shahidi, Rusak, Manning, Fradkin, Horton — The Coasean Singularity (NBER); Howard Yu — Coase vs. Claude

    Contradicting evidence: Warin — From Coase to AI Agents (California Management Review)

  21. v1.1Changed
    History of the firm

    We believe that firms will reorganise around protocols rather than disappear.

    Previous wording: Firms reshape around protocols rather than disappear.

    Stressed if agent-mediated markets consistently replace firms without retaining an accountable organisation or coordination boundary.

    Editorial hypothesis draft; existing evidence carried forward, not a new research run.

    Supporting evidence: The Headless Firm (arXiv 2602.21401); Howard Yu — Coase vs. Claude

    Contradicting evidence: None recorded.

  22. v1.1Changed
    Agency

    We believe that agency requires comparing possible futures before acting.

    Previous wording: Agency is comparing possible futures before acting. Reacting to the present is control.

    Stressed if systems without counterfactual comparison demonstrate the same adaptive, purposeful choices as systems that compare futures.

    Editorial hypothesis draft; existing evidence carried forward, not a new research run.

    Supporting evidence: Russell & Norvig — AI: A Modern Approach; Kasirzadeh & Gabriel — Characterizing AI Agents for Alignment and Governance

    Contradicting evidence: None recorded.

  23. v1.1Changed
    Agency

    We believe that an agent’s permitted autonomy is set by its deployer, separately from its capability.

    Previous wording: Autonomy is set by whoever deploys the agent, separately from what the agent can do.

    Stressed if capability alone reliably determines permitted actions regardless of deployment permissions and oversight.

    Editorial hypothesis draft; existing evidence carried forward, not a new research run.

    Supporting evidence: Feng, McDonald, Zhang — Levels of Autonomy for AI Agents; Shavit et al. (OpenAI) — Practices for Governing Agentic AI Systems

    Contradicting evidence: None recorded.

  24. v1.1Changed
    Agency

    We believe that fixed workflows should be the starting point, with autonomy added where the task needs it.

    Previous wording: Start with fixed workflows. Add autonomy only where the task needs it.

    Stressed if autonomous systems consistently outperform fixed workflows on predictable tasks after cost, reliability and oversight are counted.

    Editorial hypothesis draft; existing evidence carried forward, not a new research run.

    Supporting evidence: Anthropic — Building Effective Agents

    Contradicting evidence: None recorded.

  25. v1.1Changed
    The business as a state machine

    We believe that a state transition is a useful unit of work to delegate to an agent.

    Previous wording: The transition is the unit you hand to agents.

    Stressed if transition-based delegation cannot specify or verify outcomes as reliably as another unit on comparable business processes.

    Editorial hypothesis draft; existing evidence carried forward, not a new research run.

    Supporting evidence: Harel — Statecharts; LangGraph — agents as state-machine graphs

    Contradicting evidence: None recorded.

  26. v1.1Changed
    The business as a state machine

    We believe that an executable process map can keep operations and their description aligned.

    Previous wording: A process map that agents execute replaces one that only describes.

    Stressed if executed processes drift from their map as often as processes maintained with a separate descriptive map.

    Editorial hypothesis draft; existing evidence carried forward, not a new research run.

    Supporting evidence: OMG — BPMN 2.0; LangGraph — agents as state-machine graphs

    Contradicting evidence: None recorded.

  27. v1.1Changed
    The business as a state machine

    We believe that long-running transitions require durable execution beyond saved checkpoints.

    Previous wording: Long-running transitions need durable execution, not checkpoints.

    Stressed if checkpoint-only systems recover from crashes and interrupted side effects with the same correctness as durable execution.

    Editorial hypothesis draft; existing evidence carried forward, not a new research run.

    Supporting evidence: Temporal — LangGraph plugin adds durable execution; Temporal for AI agents — durable execution guide; LangGraph inside Temporal — architecture pattern

    Contradicting evidence: None recorded.

  28. v1.1Changed
    Ontology

    We believe that agents act more reliably through a business ontology than directly on raw data.

    Previous wording: Agents should act through the ontology, not on raw data.

    Stressed if raw-data agents match ontology-based agents on business accuracy and permission enforcement at comparable cost.

    Editorial hypothesis draft; existing evidence carried forward, not a new research run.

    Supporting evidence: Palantir Ontology + AIP; Galaxy — 2026 Enterprise Ontology Playbook (Fabric IQ); Snowflake — The Agent Context Layer

    Contradicting evidence: None recorded.

  29. v1.1Changed
    Ontology

    We believe that a business ontology needs to detect its own drift as the business changes.

    Previous wording: A static model of the business drifts within weeks, so the ontology must detect drift itself.

    Stressed if static ontologies remain accurate through business changes, or automated drift checks add no value over manual review.

    Editorial hypothesis draft; existing evidence carried forward, not a new research run.

    Supporting evidence: Atlan — Active Ontology; Ken Huang — Why Ontology Matters for Agentic AI

    Contradicting evidence: None recorded.

  30. v1.1Changed
    Ontology

    We believe that every rule in the ontology needs a named owner.

    Previous wording: Every rule in the ontology has a named owner.

    Stressed if rules without named owners are resolved, corrected and enforced as reliably as rules with accountable owners.

    Editorial hypothesis draft; existing evidence carried forward, not a new research run.

    Supporting evidence: Neo4j — The knowledge layer for enterprise AI; Authority Resolution Framework (arXiv 2608.15832)

    Contradicting evidence: None recorded.

  31. v1.1Changed
    The fact layer

    We believe that agents should use approved facts rather than conversational memory for business claims.

    Previous wording: Agents work from approved facts, not from memory.

    Stressed if memory-based agents match approved-fact agents on accuracy, provenance and policy compliance.

    Editorial hypothesis draft; existing evidence carried forward, not a new research run.

    Supporting evidence: Noan — fact infrastructure for agentic companies; Neo4j — The knowledge layer for enterprise AI

    Contradicting evidence: None recorded.

  32. v1.1Changed
    The fact layer

    We believe that one approved fact change should reach every agent using it without separate manual updates.

    Previous wording: A fact changed once reaches every agent at once.

    Stressed if dependent agents keep using superseded facts after the update, or coordinated manual updates prove equally reliable at scale.

    Editorial hypothesis draft; existing evidence carried forward, not a new research run.

    Supporting evidence: Noan — fact infrastructure for agentic companies; Atlan — Active Ontology

    Contradicting evidence: None recorded.

  33. v1.1Changed
    The fact layer

    We believe that changes to the business model should trigger a rebuild of the whole ontology graph.

    Previous wording: When the model of the business changes, rebuild the whole graph.

    Stressed if incremental updates preserve consistency and agent correctness as reliably as a full rebuild at lower cost.

    Editorial hypothesis draft; existing evidence carried forward, not a new research run.

    Supporting evidence: Ken Huang — Why Ontology Matters for Agentic AI

    Contradicting evidence: None recorded.

  34. v1.1Changed
    Learning organisation

    We believe that a business learns when observed outcomes can overturn its recorded facts and rules.

    Previous wording: A business learns when outcomes can overturn its facts.

    Stressed if recording contradictions and revising facts and rules produces no improvement in subsequent decisions or outcomes.

    Editorial hypothesis draft; existing evidence carried forward, not a new research run.

    Supporting evidence: Argyris — Double-loop learning; Senge — The Fifth Discipline

    Contradicting evidence: None recorded.

  35. v1.1Changed
    Learning organisation

    We believe that agents can propose useful changes to their skills and workflows for people to review.

    Previous wording: Agents propose changes to their own skills and workflows; people accept them.

    Stressed if agent-proposed changes fail to improve measured outcomes or create more review work than human-authored changes.

    Editorial hypothesis draft; existing evidence carried forward, not a new research run.

    Supporting evidence: Indigo.ai — Self-improving agents

    Contradicting evidence: None recorded.

  36. v1.1Changed
    Learning organisation

    We believe that business learning can live in facts and rules without changing model weights.

    Previous wording: Learning lives in facts and rules, not in model weights.

    Stressed if weight updates or learned memory are necessary to achieve outcomes that revised facts and rules cannot match.

    Recent work argues agents should also learn in their weights, and that learned memory beats designed memory. Editorial hypothesis draft; existing evidence carried forward, not a new research run.

    Supporting evidence: Survey of Agent Memory (arXiv 2602.06052)

    Contradicting evidence: Position: Modular Memory is the Key to Continual Learning Agents (ICML); Xiong, Hu, Clune — ALMA: meta-learning memory designs

  37. v1.0New
    Why a living textbook

    First published.

  38. v1.0Holds
    Why a living textbook

    A static book about agents goes out of date within one capability doubling.

    H1: evidence snapshot recorded.

    Supporting evidence: METR — Time Horizons

    Contradicting evidence: None recorded.

  39. v1.0Holds
    Why a living textbook

    Updates can be automated. The decision to publish them cannot.

    The counterpoint source supports this hypothesis: quality assurance is the part automation has not solved.

    Supporting evidence: ISPOR 2026 — GenAI framework for living reviews; Song et al. — Phases of Living Evidence Synthesis Using AI (JMIR); ISPOR 2026 — Living literature reviews: maintaining quality

    Contradicting evidence: None recorded.

  40. v1.0Holds
    Why a living textbook

    Publish what stopped being true, not only what is new.

    H3: evidence snapshot recorded.

    Supporting evidence: Elliott et al. — Living systematic reviews; CASRAI — Definition: living review

    Contradicting evidence: None recorded.

  41. v1.0New
    History of the firm

    First published.

  42. v1.0Holds
    History of the firm

    Firms exist because using the market costs something.

    H1: evidence snapshot recorded.

    Supporting evidence: Coase — The Nature of the Firm; Williamson — Markets and Hierarchies

    Contradicting evidence: None recorded.

  43. v1.0Stressed
    History of the firm

    Agents cut the cost of searching, negotiating and enforcing, so more work can move outside the firm.

    Platform dependence can raise the external costs that agents lower inside the firm.

    Supporting evidence: Shahidi, Rusak, Manning, Fradkin, Horton — The Coasean Singularity (NBER); Howard Yu — Coase vs. Claude

    Contradicting evidence: Warin — From Coase to AI Agents (California Management Review)

  44. v1.0Holds
    History of the firm

    Firms reshape around protocols rather than disappear.

    H3: evidence snapshot recorded.

    Supporting evidence: The Headless Firm (arXiv 2602.21401); Howard Yu — Coase vs. Claude

    Contradicting evidence: None recorded.

  45. v1.0New
    Agency

    First published.

  46. v1.0Holds
    Agency

    Agency is comparing possible futures before acting. Reacting to the present is control.

    H1: evidence snapshot recorded.

    Supporting evidence: Russell & Norvig — AI: A Modern Approach; Kasirzadeh & Gabriel — Characterizing AI Agents for Alignment and Governance

    Contradicting evidence: None recorded.

  47. v1.0Holds
    Agency

    Autonomy is set by whoever deploys the agent, separately from what the agent can do.

    H2: evidence snapshot recorded.

    Supporting evidence: Feng, McDonald, Zhang — Levels of Autonomy for AI Agents; Shavit et al. (OpenAI) — Practices for Governing Agentic AI Systems

    Contradicting evidence: None recorded.

  48. v1.0Holds
    Agency

    Start with fixed workflows. Add autonomy only where the task needs it.

    H3: evidence snapshot recorded.

    Supporting evidence: Anthropic — Building Effective Agents

    Contradicting evidence: None recorded.

  49. v1.0New
  50. v1.0Holds
    The business as a state machine

    The transition is the unit you hand to agents.

    H1: evidence snapshot recorded.

    Supporting evidence: Harel — Statecharts; LangGraph — agents as state-machine graphs

    Contradicting evidence: None recorded.

  51. v1.0Holds
    The business as a state machine

    A process map that agents execute replaces one that only describes.

    H2: evidence snapshot recorded.

    Supporting evidence: OMG — BPMN 2.0; LangGraph — agents as state-machine graphs

    Contradicting evidence: None recorded.

  52. v1.0Holds
    The business as a state machine

    Long-running transitions need durable execution, not checkpoints.

    H3: evidence snapshot recorded.

    Supporting evidence: Temporal — LangGraph plugin adds durable execution; Temporal for AI agents — durable execution guide; LangGraph inside Temporal — architecture pattern

    Contradicting evidence: None recorded.

  53. v1.0New
    Ontology

    First published.

  54. v1.0Holds
    Ontology

    Agents should act through the ontology, not on raw data.

    H1: evidence snapshot recorded.

    Supporting evidence: Palantir Ontology + AIP; Galaxy — 2026 Enterprise Ontology Playbook (Fabric IQ); Snowflake — The Agent Context Layer

    Contradicting evidence: None recorded.

  55. v1.0Holds
    Ontology

    A static model of the business drifts within weeks, so the ontology must detect drift itself.

    H2: evidence snapshot recorded.

    Supporting evidence: Atlan — Active Ontology; Ken Huang — Why Ontology Matters for Agentic AI

    Contradicting evidence: None recorded.

  56. v1.0Holds
    Ontology

    Every rule in the ontology has a named owner.

    H3: evidence snapshot recorded.

    Supporting evidence: Neo4j — The knowledge layer for enterprise AI; Authority Resolution Framework (arXiv 2608.15832)

    Contradicting evidence: None recorded.

  57. v1.0New
    The fact layer

    First published.

  58. v1.0Holds
    The fact layer

    Agents work from approved facts, not from memory.

    H1: evidence snapshot recorded.

    Supporting evidence: Noan — fact infrastructure for agentic companies; Neo4j — The knowledge layer for enterprise AI

    Contradicting evidence: None recorded.

  59. v1.0Holds
    The fact layer

    A fact changed once reaches every agent at once.

    H2: evidence snapshot recorded.

    Supporting evidence: Noan — fact infrastructure for agentic companies; Atlan — Active Ontology

    Contradicting evidence: None recorded.

  60. v1.0Holds
    The fact layer

    When the model of the business changes, rebuild the whole graph.

    H3: evidence snapshot recorded.

    Supporting evidence: Ken Huang — Why Ontology Matters for Agentic AI

    Contradicting evidence: None recorded.

  61. v1.0New
    Learning organisation

    First published.

  62. v1.0Holds
    Learning organisation

    A business learns when outcomes can overturn its facts.

    H1: evidence snapshot recorded.

    Supporting evidence: Argyris — Double-loop learning; Senge — The Fifth Discipline

    Contradicting evidence: None recorded.

  63. v1.0Holds
    Learning organisation

    Agents propose changes to their own skills and workflows; people accept them.

    H2: evidence snapshot recorded.

    Supporting evidence: Indigo.ai — Self-improving agents

    Contradicting evidence: None recorded.

  64. v1.0Stressed
    Learning organisation

    Learning lives in facts and rules, not in model weights.

    Recent work argues agents should also learn in their weights, and that learned memory beats designed memory.

    Supporting evidence: Survey of Agent Memory (arXiv 2602.06052)

    Contradicting evidence: Position: Modular Memory is the Key to Continual Learning Agents (ICML); Xiong, Hu, Clune — ALMA: meta-learning memory designs

The Hypothesis Diff · weekly

Know when a hypothesis is stressed or changes.

One email a week: which hypotheses hold, which are stressed, which changed, and the tools worth trying. Nobody else publishes what stopped being true.