Agent Factory

An agent-driven software factory covering the full SDLC — from idea to production and continuous improvement. Distilled from a production factory that ran multi-lane parallel builds, landing trains, and PO-governed decision loops for months; everything repo-specific has been removed, everything deterministic (gates, skills, CI, harness scripts) is included verbatim.

Installation is agent-driven: point an LLM agent (Claude Code, codex or similar) at INSTALL.md inside your target repo. The agent parameterizes the deterministic pieces against your language, build chain and domain.

The three pillars

Pillar	Question it answers	Contents
planning/	Define why, what, how	Board protocol (GitHub Projects as planning truth), spec discipline, ADR + number registry, backlog discipline with closing evidence, the standing lane brief
verification/	Prove that it works	24 deterministic gate scripts (one-way ratchets, secret scanning with a working-tree leg, traceability, planning-doc teeth), the verify portfolio, falsification norms, CI in the pinned/secret-gated regime, lander duties for landing trains
interpretation/	Understand how it really works	The choices ledger (every decision an agent made on your behalf, audited per handback), memory conventions that outlive sessions, evaluation practice (persona loops, state parity, honest metrics), investigation practice (instrument-first), the continuous-improvement loop

Shared infrastructure: skills/ (35 vendored, hash-locked agent skills), harness/ (lane launcher, worktree ritual, board bootstrap, report schema), user-level/ (what goes into ~/.claude so the factory works from any checkout).

The core loop (one screen)

idea → board item (Planned, with a full spec) → PO approval
  → lane brief from the standing template → parallel lane in its own worktree
    (agent authors; the wrapper commits; red/green falsification)
  → handback → CHOICES AUDIT (ledger entry per invented decision)
  → landing train: --no-ff pinned SHAs → cross-checks → cheap gates
    → FULL local verify green → push main → CI deploys
  → board sweep (Done + archive) → evaluation → findings → new board items
  → memory: lessons that survive the session

Three non-negotiables inherited from the source factory:

1. Falsify the apparatus before the product — a gate or test that has never been seen red proves nothing.
2. Never weaken a criterion to pass — ratchets move one way; baselines are updated only through each gate's own --update mechanism.
3. Honesty on every surface — unavailable is never rendered as empty; advisory output is labeled advisory.