Changelog
This page tracks meaningful updates to the manual. Smaller copy-edits and SEO tweaks are not listed; the footer shows the last updated date.
2026-07-06 - SVG diagrams and localized figure bodies
Three figures graduate from card layouts to theme-aware SVG: the six-phase loop now draws its real topology (six nodes, arrows, and the two failure routes back to Plan), the five governance layers render as a defense-in-depth stack, and the 90-day arc becomes a timeline with role chips. All five figures now render localized bodies in the Romanian edition (previously only captions were translated), and the arc figure's stale CLAUDE.md labels were corrected to AGENTS.md to match the text. Presentation only; the figures' content is unchanged.
2026-07-05 - How to review an agent diff (Chapter 5 + Appendix B.8)
Chapter 5's review phase gains its human half, inserted after the two agent reviewers and before verify - the read only a human can do, which the book had invoked as the human floor without ever teaching. It starts from why an agent diff fails in kind, not in degree: a human diff fails at execution and you hunt the mistake line by line, while an agent diff compiles, passes lint, and reads idiomatic, failing instead at intent and context. So the read order inverts into five steps - the diff-stat against the plan before any code, the tests read for what they assert, the boundaries the team has rules about, a grep of every new name the diff introduces, and only then line by line, where the minutes go to business correctness and architectural fit. The calibration line anchors it: fluency is not correctness, and an agent bug looks like the code a good engineer would write for a slightly different task. New Appendix B.8 agent-diff read order one-pager (the template count is now eight). Chapter 9's pattern four gains a closing back-pointer naming this read order as the human floor made concrete.
2026-07-05 - Context hygiene (Chapter 5 + Appendix B.7)
Chapter 5 gains a Context hygiene section, inserted after the inner/outer loop vocabulary paragraph, that teaches the daily craft the book had named but never taught - the window is the agent's working memory, and everything loaded into it competes with reasoning. It names the four signs of contamination (the agent re-answers a settled question, cites a stale version of a file it edited, forgets a constraint it honored, or degrades in edit quality late in a long session), the session-boundary discipline of one unit of work per session with durable state read back from the repository, and compaction as a handoff rather than a continuation, where anything not written to a file by then is gone. New Appendix B.7 context-hygiene one-pager (the template count is now seven). Chapter 1's context-window section gains a one-sentence forward pointer to the new section.
2026-07-05 - Appendix A worked example
Appendix A gains a worked pass of its own rubric, closing the one gap in a book built on worked examples: the cost section had no numbers. The new subsection runs the Chapter 10 manager sidebar's 20-engineer financial-services engagement through the appendix - 13 Team seats plus 7 Pro seats, the four-category TCO list, and the bounding heuristic - using the engagement's published operational numbers (41% of merged PRs agent-touched, cycle time 28% below baseline, defects flat) as the value anchor. The cost-side rates are labeled as illustrative round numbers, not quotes, consistent with the appendix's stance that specific prices go stale quarterly. The honest headline lands in a small table: the sticker is the smallest line, and the human-time categories dwarf it. The close holds Chapter 10's boundary: the manager defends the operational number, not a projected ROI.
2026-07-05 - Two kinds of hooks (Chapter 9 pattern three)
Chapter 9 pattern three expands from hookify rules into two kinds of hooks: the agent's pre-tool-use hook (hookify) and git's pre-commit/pre-push hook. The new material frames the git gate on the agentic spine - the agent is just another author with write access, so a deterministic commit/push gate applies to it for free, and the --no-verify escape hatch is closed by a hookify deny rule. It draws the honest line too: local hooks are fast feedback, while the same checks in CI are the gate you cannot skip from your laptop, which is pattern eight's gate the agent cannot edit. Recap, closing, and the artifact box updated to match; the pattern count stays eight.
2026-06-11 - Scripted subagent orchestration (/workflows)
Chapter 1's subagents section gains a paragraph on programmatic orchestration: Claude Code's /workflows, which has the agent write a deterministic script that dispatches subagents across phases, parallel/pipeline fan-out, and schema-validated handoffs, with a verification gate. Framed on-thesis - not a new primitive but harness convenience over the subagent primitive, where the determinism is the only real difference from model-driven parallel dispatch (e.g. a Superpowers-style plugin). Dated to mid-2026.
2026-06-10 - The outer loop (pattern eight)
Chapter 9 grows an eighth pattern covering the outer-loop trend - re-invoking the agent on an interval, a schedule, or a queue until a condition holds. The pattern traces the lineage (the 2023 self-grading loops; Ralph Wiggum in mid-2025; the late-2025 vendor wave of background and scheduled agents; the loop as a first-class surface by spring 2026) and installs the controls: a five-line loop contract, the loop-eligibility test, four loop kill signals, and the morning review floor. Chapter 5 gains the inner/outer loop vocabulary paragraph. New Appendix B.6 outer-loop contract one-pager (the template count is now six). Five new Appendix C entries under a new outer-loop and autonomy sources group. Dated-claims note bumped to June 2026.
2026-05-27 - Permissions / Sandbox primitive
Permissions / Sandbox promoted to a named primitive - the third slot in the inventory, after Context window and Tools. Two halves like Memory: the agent-level decision layer (Allow / Ask / Deny + auto mode) and OS-level enforcement (Seatbelt / bubblewrap / restricted tokens / WSL2). The vocabulary note in Chapter 1 was rewritten to name the convergence test that promoted this primitive while leaving telemetry as a control layer. Chapter 3 gains a framing paragraph binding three of its five layers (permissions, hooks, sandbox) to the new primitive's configuration surfaces; the five-layer defense-in-depth narrative is preserved unchanged. Inspection-points count in Chapter 2 dropped from nine to eight (the two halves of P/S collapsed to one). Diagram updated to 8 cells. Three new Appendix C entries source the Claude Code / Codex / opencode implementations.
2026-05-27 - Changelog + last-updated footer
Added this page and a "Last updated <date>" line in the footer, plus a Changelog link in the contact row. The intent: returning readers can see what moved since their last visit without diffing chapters.
2026-05-27 - Memory primitive + open-set framing
Memory promoted to a named primitive across the book; Chapter 1 retitled "The primitives" and the structural argument rewritten as "named primitives + the recursive primitive (subagents)". Dropped the closed "six primitives" count throughout; the list is now framed as open-set. New Memory section covers two halves - manually defined memory (AGENTS.md/CLAUDE.md, agent-agnostic) and the auto-memory system (Auto Memory, Auto Dream - currently Claude-Code-led). Diagram updated. Chapter 6 gains a one-paragraph framing intro anchoring AGENTS.md as the team-shareable memory layer. Three new Appendix C entries source the Memory claims.
2026-05-27 - SEO pass: per-chapter URLs
Every chapter and appendix now has its own URL (20 pages total). Old /#chapter-N bookmarks redirect. Each page got a unique title, meta description, canonical URL, prev/next nav, and TechArticle + BreadcrumbList structured data. Hero gained a three-button CTA row. Schema upgraded from Article to Book + Organization + FAQPage. Cross-section anchor rewriting, AGENTS.md de-linking, /read/ mode for the single-page reading experience. Built-in 404 page, llms.txt for AI answer engines, cover.webp.
2026-05-26 - Feedback-pass polish
External-reviewer pass: TOC chapter-to-part mismatch fixed, figure numbering dropped (web-manual style), new Source-note and Artifact callout components with light + dark variants, foreword bio trimmed to four sentences and full version moved to a new About-the-author section, hero gained a control-thesis dek, AGENTS.md links collapsed to at most one per chapter, per-section ¶ copy-link anchors, callout stack tightening.
2026-05-26 - First public version
Manual published at ship-it-with.ai. Ten chapters across three parts (Architecture, Method, Reality), three appendices, plus foreword, prologue, closing.