The AI Software Factory Guide

Build the system around the agent

A field guide to designing, building, proving, operating, and improving an AI Software Factory: an engineering system in which humans define intent and accept risk, bounded agents plan, implement, validate, and recover, and independent evidence — not an agent saying "done" — decides what advances.

Intent → Plan → Define Agent → Execute through Harness → Apply Skills → Evaluate → Improve → Deliver Software

Read it front to back, or enter at the part that matches your question.

The six parts

Part	Question it answers	Chapters
I. Understand	What is an AI Software Factory, what are its parts, and what principles hold it together?	1–3
II. Design	What records, decisions, authority, and economics must exist before any agent runs?	4–10
III. Build	How do you assemble capabilities, runtime, harnesses, environments, AI layers, and workflows?	11–26
IV. Prove	How do you know the factory's output is correct, safe, and releasable?	27–33
V. Operate	How do you run it as a production platform?	34–38
VI. Improve	How does it get better without authorizing itself?	39–44

Start with How to read this guide, then Chapter 1. The book map lists every chapter and appendix.

The system in one view

flowchart TB
    Human["Human intent, policy, and decisions"] --> Control["Control plane and orchestration"]
    Factory["Agent Factory: agents, skills, tools, profiles, and evals"] --> Control
    Control --> Contract["Frozen execution contract"]
    Contract --> Outer["Outer harness"]
    Outer --> Inner["Inner coding harness"]
    Inner --> Environment["Development environment"]
    Environment --> Compute["Compute infrastructure"]
    Inner --> Candidate["Candidate and run record"]
    Candidate --> Verify["Independent verification and evidence"]
    Verify --> Decision["Human or policy decision"]
    Decision --> Delivery["Delivery and production outcome"]
    Delivery --> Learn["Governed feedback and improvement"]
    Learn --> Factory

Loading

The downward path delegates bounded capability. The upward path reports observations, evidence, and outcomes. An executor cannot grant itself authority or certify its own material work. Chapter 2 draws the whole system.

Three definitions to retain

Agent Factory creates, versions, evaluates, publishes, and governs reusable capabilities such as agents, skills, tools, model profiles, and configurations.

AI Software Factory composes people, policy, capabilities, execution, verification, delivery, and feedback from intent through validated production value.

Mission Control is the living implementation and case study for the control-plane responsibilities required to govern execution, evidence, and human authority. It is not the definition of the complete factory.

Governing principles

1. Humans own intent, judgment, material risk, and irreversible decisions.
2. Agents operate only inside explicitly granted authority.
3. Independent evidence—not an agent saying "done"—determines readiness.
4. Simple deterministic work should remain deterministic.
5. Failure must be detectable, bounded, recoverable, and attributable.
6. Autonomy increases only when measured outcomes justify it.
7. Learning may be automated; promotion remains governed.

Repository layout

• guide/ — the book: front matter, six parts (44 chapters), appendices (glossary, Mission Control case studies, research canon, coverage and maturity, changelog, reviewer guide, architecture communication, operator surfaces).
• archive/guide-v1/ — the previous curriculum, preserved unchanged for provenance. docs/plans/coverage-map.md records where every v1 chapter went.
• source-material/ — Jay's original mission, study guide, and capability taxonomy, preserved verbatim.
• site/ — the documentation site (Guide · Atlas · Reference · Glossary).
• docs/ — plans, reviews, usability notes, and the nightly backlog.

Conventions

Each chapter follows the same shape: the problem, how it works, how to build it, failure modes, an honest "In Mission Control" note pinned to a studied commit, "Retain this", and "Go deeper". Terms are bolded on first use. Diagrams are Mermaid. Marked infographic slots hold the place for Jay's own graphics; a first-party diagram carries the concept until then.

Every "In Mission Control" section separates implemented, partial, and future capability. Documentation breadth is not proof of an operational implementation; see Coverage and maturity.

Site

cd site && npm install && npm run dev

npm run build, npm test, npm run lint, and npm run links must pass before publishing.

Shared FDLC navigation

The global header consumes site/lib/global-navigation.generated.ts, generated from FDLC's canonical lib/global-navigation.ts. Run node site/scripts/sync-global-navigation.mjs --source=/path/to/FDLC after a navigation change, and add --check to verify against the exact qualified FDLC source before releasing both applications. Do not edit the generated copy. Guide-specific active states remain local.