What this is

A governance layer for teams whose AI agents propose, build, and ship real product work -- not a specific app, but a set of concrete workflows, state machines, and file formats: adversarial review before implementation, a hard cost circuit breaker, a quality-gate engine, and an evidence-based learning system that stops agents from self-certifying unverified improvements.

The problem this solves

Teams running Claude Code, Codex, or similar agents on real work run into the same failure modes repeatedly: an agent proposes something and immediately starts building with no adversarial check, a multi-agent debate silently burns an unbounded amount of tokens, work gets called "done" without a clear definition of what that means, and worst of all -- an agent tries something once, it seems to work, and that becomes the new unquestioned default with zero reproducible proof.

What's included

• Adversarial blue-team/red-team review workflow with a fixed decision-brief format
• Hard cost circuit breaker: per-round word caps, max rounds, per-review and monthly dollar ceilings, automatic deadlock reports
• Quality-gate engine: fixed gate categories, any failing gate blocks promotion, evidence-path requirement (not just a claim)
• Evidence-based learning state machine (unverified -> candidate -> verified -> promoted, with blocked/deprecated branches)
• The exact fields a recommendation record needs to count as a real, checked improvement
• Skill-library-bridge pattern for keeping agent prompts lean without sacrificing depth
• English and Traditional Chinese versions, both fully written out (not machine-translated from each other)

FAQ

Do I need a specific multi-agent framework to use this?
No -- this is workflow and state-machine methodology, applicable whether you're running Claude Code, Codex, a custom agent orchestration, or even a human review process you want to make more rigorous.

Is this only for large teams?
No -- a solo developer using AI agents heavily benefits from the same discipline: the cost circuit breaker and the evidence-based learning system both prevent silent failure modes regardless of team size.

What if we already have code review and CI -- do we need this too?
CI/CD catches whether code works; this catches whether the decision to build it was sound, whether cost stayed bounded, and whether a claimed improvement is actually reproducible rather than a one-off fluke an agent is treating as gospel.

中文版

這是什麼

給那些讓AI agent提案、建構、上線真實產品工作的團隊用的治理層——不是特定app，而是一套具體的工作流程、狀態機與檔案格式：實作前對抗式審查、硬性成本斷路器、品質閘門引擎、防止agent自己宣稱驗證通過卻拿不出證據的學習系統。

這解決的是什麼問題

用Claude Code、Codex或類似agent做真實工作的團隊，會一直重複踩到同樣的失敗模式：agent提案完馬上就開始動手蓋，沒有任何對抗式檢查；多agent辯論悄悄燒掉無上限的token；工作被宣稱「完成」卻沒有明確的完成定義；最糟的是——agent試了一次做法，看起來成功了，這就變成沒人質疑的新預設行為，完全沒有可重現的證明。

包含什麼

• 對抗式藍軍/紅軍審查流程，附固定的決策簡報格式
• 硬性成本斷路器：每輪字數上限、最大輪數、每次審查與每月的金額上限、自動僵局報告
• 品質閘門引擎：固定關卡分類，任何關卡沒過就擋下晉級，要求證據路徑而不是空口宣稱
• 以證據為本的學習狀態機（unverified → candidate → verified → promoted，含blocked/deprecated分支）
• 一筆建議紀錄要具備哪些欄位，才算得上是真正被檢驗過的改善
• Skill庫橋接模式，讓agent提示詞保持精簡又不犧牲深度
• 中文版與英文版皆完整獨立撰寫，不是互相機器翻譯

常見問題

一定要用特定的多agent框架才能用這個嗎？
不用——這是工作流程與狀態機方法論，不管你是用Claude Code、Codex、自訂的agent協作，甚至是想讓現有人工審查流程更嚴謹，都適用。

這只適合大團隊嗎？
不是——就算是一個人大量使用AI agent，也一樣受益於同樣的紀律：成本斷路器跟證據為本的學習系統，不管團隊規模大小，防的都是同一種悄悄發生的失敗模式。

我們已經有code review跟CI了，還需要這個嗎？
CI/CD檢查的是「程式碼有沒有跑起來」；這個檢查的是「當初決定要蓋這個東西的判斷合不合理」「成本有沒有守住上限」「宣稱的改善是不是真的可重現，而不是agent把一次僥倖成功當成金科玉律」。