Revision History

This handbook is a living document, reviewed monthly with fast patches for major ecosystem changes. This page is its activity log. If the systems we describe keep an audit trail, so should we.

July 2026 Edition

The launch edition. A full editorial pass across both tracks, plus new logged material from the June operator runs.

• New logged evidence throughout. The skill audit (ch 6) now includes a real audit of a freshly provisioned FlowWink instance. The vendor chapter (ch 12) carries the logged manage_leads silent-failure case. The mandate layer (ch 14) gained a third quiet proof — the empty support queue. The enterprise chapter (ch 11) gained “Four Days, Logged”: one operator running a business cycle from zero to 860,000 SEK ARR in four days.
• The MCP adoption spectrum (ch 11): how vendors are opening up — early movers, fast followers, and the community wrappers that queue outside the doors of the laggards. Plus the author’s reflection on digital sovereignty and loop engineering.
• The June 2026 model shutdowns and what they mean for operator architecture (ch 7).
• Mid-2026 market scale signals (ch 11): the Gartner $206.5B forecast, SAP’s “Autonomous Enterprise,” and vendor adoption figures.
• The value thread made explicit (chs 9, 11, 15): surfaced vs realized value, the operator-value equation, and the receiving organization — anchored in Per Clingweld’s Almedalen takeaway and Satya Nadella’s “token capital” (ch 8, 13).
• The knowledge question (ch 6): Google’s Open Knowledge Format and knowledge hygiene — what the agent reaches must also be understandable.
• Author’s note added — Why I Built This (foreword).
• Skill counts updated to FlowWink’s current surface: 300+ skills across 60+ modules.
• The Cursor acquisition (Builder Edition, foreword + ch 5): SpaceX’s $60B purchase of Anysphere — the control-plane thesis got its price tag. Model-name references refreshed to the current generation.
• BCG AI at Work 2026 woven in (chs 11, 15): 30% of organizations now run agents in production; clarity beats tools by twenty points; the evaporating-workday finding as population-scale proof of the value equation.
• All four payment giants (ch 13): Mastercard and Visa joined Google and Stripe/OpenAI with shipped agentic-payment programs.
• Agent cost governance (chs 11, 14): the mid-2026 enterprise cost-capping wave (“an agent is a meter that runs”), and the budget boundary as the mandate component most deployments forget.
• The Fixed-Cost Workhorse (ch 11): open-weight workhorse models on owned or partner-leased GPU rigs — the excavator analogy, the night-shift-to-kanban rhythm, and the CFO principle “fixed cost where the volume is, variable cost where the value is.”
• “Eighteen Months In” (ch 15): the vision tableau — a Tuesday at the operator-run workplace, assembled entirely from mechanisms proven earlier in the book. The chapter’s opening question finally gets its painted answer.
• Process catalog appendix: the thirteen standard business processes, what an agent runs in each today, and which lever each pulls — snapshot of the reference implementation’s living coverage map, with the funnel out to the full catalog and its gap lists.
• End-of-book survey (Business ch 15 and Builder “Closing Words”): three taps at the finish line — NPS, interest in handbook three, and what it should prioritize. The readers are now research input for the book being researched. Followed by one optional free-text note — a name, a few lines, and a checkbox for whether it can be used publicly as a reference. Entirely skippable; nothing is published without that checkbox. Both editions now close this way.
• Chapter feedback. One-tap thumbs at the end of every chapter — anonymous, once per chapter. Reader feedback becomes editorial findings; the monthly edition is the action.
• Listen buttons on every chapter. Browser-native text-to-speech (Web Speech API): play/pause/resume, speed control, paragraph highlighting, estimated listening time. Zero cost, zero audio files to regenerate — the living book reads its current text aloud.
• The handbook became agent-readable. Its own MCP surface (Streamable HTTP at /api/mcp — list, read, and search every chapter), plus llms.txt and raw markdown per chapter. The book about MCP now speaks MCP.
• “Running the Vendor Conversation” (ch 12): the leadership playbook for vendor pressure — where to press first (CRM and finance before dashboard-heavy ERP), the answer decoder (“no demand” → count the wrappers), the 90-day/two-quarter/twelve-month clock, and the capability-lag vs strategic-closure switching rule.
• “The Two Dimensions” map (ch 11): vertical agents vs horizontal operators × rule-work vs judgment-work — the quadrant picture of where agentic AI pays off fastest, with triage named as the first job. Vertical/horizontal vocabulary bridged in ch 10.
• “The Post-SaaS Question” (ch 13): the reusable-rocket first-principles frame, Nadella’s “SaaS is dead” anchor, the laggard inversion (cloud stragglers may skip the SaaS era entirely), and the renewal-as-a-bet leadership moves — explicitly hypothesis-tagged.
• “What Is AI Actually Expected to Contribute?” (ch 11): the four expectations ranked by evidence — cost-cutting last, time saved evaporates without redirection, revenue strongest in the log, decision quality as the meta-contribution. The forest answer: attention redistributed, signal-to-decision latency compressed.
• The trust curve and the warm negotiator (ch 14): Capgemini’s trust-fell-as-adoption-tripled finding, and MIT’s tournament showing warm agents out-negotiate ruthless ones.
• The April platform cutoff (ch 7): Anthropic’s subscription-policy change as the commercial twin of June’s export-control lesson.
• Meta-Harness (Builder ch 35): the Stanford/MIT/KRAFTON paper — empirical proof that the harness matters as much as the model.
• The landscape completed (Builder ch 5): OpenAI Codex and Google Antigravity join the control-plane table — all three frontier labs now ship their own agent layer, which settles the chapter’s central claim.
• Structural fixes: The Contract Layer chapter completed and moved to follow The Agent’s Superpowers; chapter titles and operator naming made consistent throughout.
• “The Labs Came From the Other Side” (ch 7): Anthropic’s Claude Code/Cowork, OpenAI’s Codex, Google’s Antigravity — and xAI, the lab that bought its agent layer with SpaceX’s Cursor acquisition. Plus Microsoft as the proof of the trajectory: agentic mode already shipping inside Word, Excel, and PowerPoint. Why the convergence doubles the operator decision rather than replacing it — the personal agent surface vs. the standing business operator, the two budget lines a CFO should keep apart — and the direct answer to “if the giants all ship agents, why does this book run on OpenClaw?”
• The findings motto (ch 15): “the day it finds ten problems is a better day than the day it finds none — the problems were there yesterday too.” The leadership version of the drift frame, added where the evidence-tag discipline turns into daily practice.
• Case-consistency pass (chs 3, 4, 11, 14): every scenario customer traced across all chapters. The blind-test anomalies now carry their values (€45,000 won/unsigned, €22,000/yr expiring) where chapter three describes them; the enterprise chapter’s CFO argument references the blind test correctly instead of misattributing it; the stressed morning is explicitly framed as a re-seeded stress test; the Lindvall Systems thread is one story told across two chapters (sweep finds it, a human dispatches, four minutes resolves it) instead of two conflicting versions; and the two dunning stories (chs 10, 14) now explicitly contrast — escalation without a rule vs. a quiet stop with one.

June 2026 Edition

• Live Proof extended with Act III (the stressed morning) and Act IV (the long game — the 18.3x invoice cascade, tracked across seven weeks).
• ClawStack renamed ClawClass across the handbook, matching the project’s rename.
• Live Proof skill counts synced with the platform’s expanding MCP surface.

May 2026 — First Edition

• Initial publication: sixteen chapters plus appendices. The April 19 detection proof, the operation proofs, the architecture decisions, the failure modes, the mandate layer, and the ninety-day path.

Every change lands in the public repository with a commit history — the full audit trail is at github.com/magnusfroste.