Garry Tan — “Building with AI” 통합 아키텍처

7개 시리즈 종합 분석 + Drewgent 보완 제안

Author: Garry Tan (Y Combinator CEO) Series: 7개 기사 (2026년 4월~5월) Target: Drewgent Architecture Review Status: HP-2/HP-3 구현 완료 (2026-05-15)


시리즈 요약

#제목날짜조회핵심 주장
1Thin Harness, Fat SkillsApr 11153만5개 정의 소개
2Resolvers: The Routing TableApr 1632만Context routing
3On the LOC ControversyApr 1813만AI 생산성 측정
4Naked Models Are StupiderApr 1915만모델=엔진,≠자동차
5The Skillify ManifestoApr 22?LangChain 비판
6Meta-Meta-PromptingMay 5140만Personal AI OS
7The Complexity RatchetMay 1216만90% 테스트 커버리지

Garry Architecture 핵심 정리

5개 정의 (Article 1)

Definition 1: Skill File Markdown 절차 — WHAT이 아닌 HOW를 가르침. 파라미터로 호출.

Definition 2: Harness LLM 실행 프로그램 — 4가지만 함: 모델 루프, 파일读写, 컨텍스트 관리, 안전 강제. Anti-pattern: fat harness (40+ tools) + thin skills.

Definition 3: Resolver “task type X → document Y” 매핑 테이블. on-demand doc load.

Definition 4: Latent vs. Deterministic Latent: 모델 판단/합성. Deterministic: Same input → Same output.

Definition 5: Diarization 50개 문서 → 1 page judgment 출력. RAG/SQL 불가.

The Complexity Ratchet (Article 7)

90% coverage → no bugs → safe refactoring → faster feature dev → more coverage
→ 시스템 개선만 가능, 나빠질 수 없음 (棘輪)

~970,000 lines, 665 test files, 72小时内 14 PRs merged.


Drewgent 매핑

항목Garry ArchitectureDrewgent 현재상태
Fat Skillsmarkdown procedures~/.drewgent/skills/*.md✅ 완료
Thin Harness~200 lines CLIrun_agent.py (8.7K줄)⚠️ 분해 필요
HP-2: Latent DetectionLatent vs Deterministic 구분_is_latent_task() 추가✅ 완료
HP-3: Complexity Ratchet90% test coverage3-phase QA gate + delivery blocking✅ 완료
Context Resolveron-demand doc loadRESOLVER.md 생성 완료✅ 완료

HP-3 3-Phase QA Gate 구현 (2026-05-15)

Flow

Latent task 감지 (implement/build/create 등)
  → _is_latent_task() = True
  → emit_qa_gate("contract") + self._qa_task_id set
  → QA_GUIDANCE_TEMPLATE injected into system prompt
      ↓
[Turn 1] contract.json 작성 (criteria)
  → emit_turn_end() → micro emitted
      ↓
[Each turn] micro-qa.json累积
      ↓
[agent-complete] emit_qa_gate("full")
  → full-qa.json must have all_criteria_met=true
      ↓
gateway: qa_gate_status check
  ├── True → delivery 통과
  └── False → ⚠️ blocked

Evidence Files (P2-hippocampus/qa-evidence/{task_id}/)

  • contract.json — acceptance criteria (written by agent before coding)
  • micro-qa.json — per-step verification (accumulated across turns)
  • full-qa.json — final verification (all_criteria_met required for unblock)

Files Modified

FileLinesChange
run_agent.py+65helpers, contract/micro/full emit, qa_gate_status in result
agent/prompt_builder.py+18QA_GUIDANCE_TEMPLATE constant
gateway/run.py+17delivery blocking check
agent/signal_processor.py~90_on_qa_gate() skeleton auto-generation
禁task_qa_gate.neuronfull rewriteHP-3 implementation documented
RESOLVER.mdupdatedHP-2 latent keywords + flow
garry-tan-complexity-ratchet-2026.mdupdatedHP-3 implementation note