Hey HN, I'm Fred. I designed this from an iPhone during a 6-week digital detox in Vietnam — my partner banned my laptop.

  Without a computer I couldn't chase the weekly model releases. I was
  stuck with a harder question: not "how to code faster with AI" but
  "how to make agentic coding reliable." I read papers on separating
  planning from execution in agent systems — GoalAct (Chen 2025) shows
  +12.2% success when you split them [1]. That became the core design.

  GAAI is a .gaai/ folder — markdown, YAML, bash. No SDK. You drop it
  into any project. It gives you two agents with a hard gate between
  them:

  - Discovery structures what to build (stories, acceptance criteria).
    Never writes code.
  - Delivery ships validated stories autonomously (plan, implement, QA).
    Never decides scope.
  - The backlog is the contract. Nothing gets built that isn't in it.
  - Memory persists across sessions — session 47 knows what session 1
    decided.

  I stress-tested it by building a full SaaS — 7 Cloudflare Workers,
  auth, billing, matching engine, analytics. First 4 days: 39 stories
  shipped, 260 tests passing, 16K LOC. Today: 176 stories, 84K LOC.
  96.9% cache read rate — agents build forward, not from scratch.

  vs. AGENTS.md: solves one session, not cross-session memory or scope
  drift. vs. LangGraph/CrewAI: code-first orchestration for building AI
  systems. GAAI governs using AI tools. Different level.

  Honest trade-offs: no programmatic enforcement (agents follow the
  files), even trivial tasks need a backlog item, freshly open-sourced.

  Install: git clone + bash install.sh --wizard. 30 seconds.

  ELv2 — use it freely in your projects. Only restriction: can't offer
  it as a competing hosted service.

  [1] https://arxiv.org/abs/2504.16563