Garry Tan — “The AI Agent Complexity Ratchet: Why 90% Test Coverage Is Required”
Series #7 — Building with AI
Meta
- Author: Garry Tan (Y Combinator CEO)
- Published: May 12, 2026
- Stats: 168K views, 768 likes, 127 reposts, 1531 bookmarks
- Series: 7th in “building with AI” series (1-6 links present)
Core Thesis
90% test coverage is required — and AI agents made it free to get.
50년간 소프트웨어 엔지니어링은 “에러 방지” 중심으로 운영됨. 에러 = 재앙이었음. 한 번의 실수가 프로덕션 크래시. 90% coverage이 해답: 시스템이 개선만 되고, 나빠질 수 없게 만드는 lock.
What He Built
Two OSS Projects
- GStack — AI coding agents better (23 specialist skills, multi-host: Claude Code, Codex, Cursor, etc.)
- GBrain — turns everything you read/write into searchable knowledge base for AI
Scale
- ~970,000 lines of code
- 665 test files
- Mostly written by Claude Code + Codex at his direction
- 15 simultaneous Conductor sessions most of the time
- 72小时内: 14 PRs merged, ~29,000 lines new code
The Complexity Ratchet
90% coverage → no bugs → safe refactoring → complexity increase possible
→ faster feature dev → more coverage → ...
システムが改善만 가능, 나빠질 수 없음 = Ratchet (棘輪)
Speed vs Quality trade-off 없음. “Ship fast, break things” vs “Move slow, ship right” — 둘 다 버림.
Key Quotes
“Speed and quality are supposed to trade off. Ship fast, break things. Move slow, ship right. Pick one.” “You don’t have to pick anymore. The unlock is 90% test coverage — and AI agents made it free to get there.” “For fifty years, that level of verification cost too much human willpower to sustain. Now the agent writes the tests alongside the code.”
Software Was Brittle (50 years)
- 에러 = 재앙 — one edge case miss = crash in production
- DB migration 위험 — bad migration = lose customer data
- 암묵적 지식 문제 — one person who understands it quits = nobody knows why it works
- 전체 시스템이 그 사람에 의존 — tribal knowledge = architectural debt
The Unlock: AI Made 90% Coverage Free
- Agent writes tests alongside code (not after)
- Human willpower no longer the bottleneck
- “The agent writes the tests alongside the code”
- Result: complexity ratchet — system can only get better, never worse
Architecture Implications for Drewgent (HP-3 Implementation: 2026-05-15)
- Test coverage as quality gate — Drewgent’s 禁task_qa_gate aligns with this thesis
- Complexity ratchet = self-improving agent — Trend Harvester fits here
- AI makes rigorous verification free — Agent writes tests alongside code
- Speed + quality no longer trade-off — parallels Drewgent’s TDD-PDCA loop
- HP-3 implemented (2026-05-15): 3-phase QA gate (contract/micro/full), delivery blocking, QA_GUIDANCE_TEMPLATE injected into system prompt for latent tasks
Related
-
GStack: https://github.com/garrytan/gstack (95K+ stars)
-
Series 1-6: Not accessible (Xitter article format requires login)
Tags
garrytan ai-agent test-coverage complexity-ratchet gstack gbrain building-with-ai yc