gstack

garry-tan‘s open-source (MIT) “AI software factory” — 23 specialized agent-skills + 8 power tools that turn Claude Code into a virtual engineering team. A flagship instance of the agentic-coding-harness, and a node that ties much of this wiki together. Tan’s framing: “the point isn’t who typed it, it’s what shipped.”

How it works

A structured sprint — Think → Plan → Build → Review → Test → Ship → Reflect — where each skill feeds the next (design docs inform reviews, test plans guide QA, shipping triggers doc updates). Roles are split across specialist agents: CEO, designer, eng manager, QA lead, security officer, release engineer.

• Commands: /office-hours, /plan-ceo-review, /review, /qa (real-browser testing + bug-fixing), /design-shotgun, /cso (security), /ship, /document-release.
• Parallel sprints via conductor — 10-15 simultaneous sessions.
• Memory via gbrain — /setup-gbrain (PGLite / Supabase / remote MCP) and /sync-gbrain keep persistent cross-session knowledge current.

Why it matters here

gstack is the convergence node of the burst: it composes agent-skills (23 of them), agent-orchestration (role-split agents), conductor (parallel sprints), and crucially gbrain (research-wiki’s knowledge base) — its Reflect step is compound-engineering (each sprint codifies learnings). It’s the clearest single example of this wiki’s agentic-coding-harness and its bridge back to research-wiki: the same author’s knowledge-base gbrain is the harness’s memory. (Caveat: README/founder claims.)

Related

garry-tan · gbrain · conductor · agent-skills · agentic-coding-harness · agent-orchestration · compound-engineering

Re-verified 2026-08-04 (staleness queue)

126.3k★ / 19.0k forks, MIT, 384 open issues and 497 open PRs — by adoption this is now among the largest artifacts in the corpus, and the open-PR count says the maintainer is the bottleneck.

Scope has grown past the 23 role-skills this page recorded: browser automation, cross-model review, GBrain knowledge-base integration, and ten AI coding platforms rather than Claude Code alone. garry-tan‘s own productivity claim is unchanged in kind and now dated — “in the last 60 days: 3 production services, 40+ shipped features, part-time, while running YC full-time” — still a first-party anecdote with no denominator, which is what this page already said.