One OS for every harness: injecting AI coding standards
카카오 AI 돛 Summit 26 | 모든 하네스에 하나의 OS를 — AI 코딩 표준 주입기
B · Useful, limited Kakao Gift server chapter lead describes a working system for injecting team coding standards into each developer's own agent harness, with operating figures; a specific, transferable design, but self-reported with no defect or review-time outcomes.
Summary based on the full transcript·Original: about 19 minutes to watch
In short
A 19-minute talk at Kakao's AI Summit 26 (published 21 October 2026) by Ma Tae-il, who leads the Kakao Gift server chapter and the Commerce AX Boost squad. With dozens of developers each running several agents, team rules on caching, retries and timeouts arrive only in PR review, after tokens and time are spent. Ma first wanted one shared harness for everyone, then dropped the idea. Instead, a tool called jansori ("nagging") inserts a minimal team operating standard into each developer's own CLAUDE.md or AGENTS.md and intervenes while the agent is writing code.
Key numbers
- 270,000+tokens in the jansori knowledge base
- ~10,000 (~4%)tokens loaded per session through on-demand routing
- 146 to 68candidate rules proposed versus adopted in the 18 September 2026 improvement round
- ~16%share of governance-relevant tasks in the last 30 days where jansori intervened
- 9skills in the shared operating standard (API, design, security, performance and others)
As reported by the source.
Korean·Video·opens youtube.com
What it reports
- A harness is described as workflow, skills, tools and context, and operating instructions around a model. A 30-hour hackathon judged on real customer payments convinced Ma that generic harnesses fail on the last mile and teams need their own.
- jansori installs with one line and a wizard for Claude Code or Codex. It writes the common standard between markers in the instruction file, leaves personal instructions untouched and syncs new versions automatically.
- The knowledge base exceeds 270,000 tokens; about 10,000 load per session via routing: common metadata and skill descriptions always, skill bodies when a task touches a governance area, references only when more evidence is needed.
- Nine skills: user, admin, server API, system design, design, event and batch processing, deployment, security, performance.
- Example: a developer asks to cache a popular-keyword list; the performance rule fires before code is written, warning that whole-response caching could show sold-out items as available. The developer narrows the cache.
- Message format: observation, consequence, question (the developer decides), then a one-line record of the choice, sent back to the jansori server.
- It stays silent unless a governance area is touched and a rule is broken; in the last 30 days it spoke on about 16% of governance-relevant tasks. (A following figure of 26% caught before review is unclear in the transcript.)
- Improvement rounds: one AI proposes rules by comparing code with rules, a second argues against them. On 18 September, 146 candidates yielded 68 rules. A "mock exam" set on Langfuse checks that it speaks when it should and stays quiet when it should; a new version ships only if it scores at least as well as the last.
- People set two things: the grading criteria and the points where the loop must stop and ask a person.
How they got it
Conference talk by the system's owner, with figures from the tool's own logs and evaluation set. Transcript from Korean auto-captions.
Limits
- No outcome data on defects, review time or rework.
- The 16% and 26% figures are not clearly reconciled in the transcript.
- Single team; technical detail deferred to a later if(kakao) talk.
Why it matters for companies
This is a practical answer to a growing governance problem: how to apply company standards to agents that every developer configures differently. Inject a small, versioned standard into each harness, load detail only when relevant, and make the agent intervene before code is written instead of after review. The rule-update loop, with an adversarial reviewer and a regression gate, is a pattern any platform or security team can reuse for AI policy content.
Key passages in translation
Korean original
다 만든 뒤에 지적보다 정하는 순간에 한 마디가 압도적으로 저렴합니다.
A single remark at the moment of deciding is far cheaper than pointing things out after everything is built.
Korean original
9월 18일 날 실제 라운드 수행에서 후보 146개 가운데 68개만 룰이 되었습니다.
In the actual round on 18 September, only 68 of 146 candidates became rules.
Korean original
AI에게 맡길 때 가장 중요한 것은 멈출 곳을 잘 지정하는 것이라고 생각합니다.
When handing work to AI, I think the most important thing is specifying well where it must stop.
Summary based on the full transcript. Read the original for the full argument and data.Watch the original on youtube.com
- Tags
- Governance, Agents.md, Evals