資料來源#
- Driving the Agent Quality Flywheel from Your Coding Agent- Google Developers Blog
- Reliability without Validity: A Systematic, Large-Scale Evaluation of LLM-as-a-Judge Models Across Agreement, Consistency, and Bias
摘要#
這項規則是:在任何改進迴圈中,提出變更的那個東西,絕不為該變更評分。 Google 的 Agent Quality Flywheel 將它表述為設計不變量:optimizer(你的 coding agent、自動化 optimizer,或你自己)提出方案;evaluation service 獨立評分——因為「會替自己工作評分的 optimizer,學到的是操弄指標,而不是改進 agent。這項小小的架構選擇,比看起來更重要。」這是以結構方式處理 古德哈特定律,而非以行為方式處理:不必寄望 optimizer 保持誠實,而是移除它取得評分的權限。
為什麼重要#
Reward hacking 通常在訓練迴圈中討論——模型操弄自己的 reward signal。同樣的動態也會在開發迴圈中運作:一個針對某項指標反覆調整 prompt、同時又負責計算該指標的 agent,最終會產出符合自身評分的結果,而不是達成使用者的目標。這種失敗很安靜,因為指標持續改善;只有獨立的 grader(或 production 流量)才會揭露偏離。解耦將「它真的變好了嗎?」從自我報告轉成外部檢查——也就是主張與測量之間的差異。
同樣的拆分在哪裡反覆出現#
wiki 已經收錄數個獨立推導出這項規則的案例,這表示它是真正的不變量,而不是某家供應商的偏好:
- Loop Engineering —— Osmani 的 maker/checker sub-agent 拆分(「maker 對自己的作業評分時過於寬容」),以及
/goal的設計:每一輪結束後由另一個模型檢查停止條件,因此撰寫程式碼的 agent 不會同時決定自己已經完成。 - LLM-as-a-Judge —— 自我評分與 judge lineage 的注意事項:與受評模型共享訓練 lineage 的 judge,會構成 validity threat;DRACO 透過 human-alignment studies 選擇 judge,並使用不重疊的 judges 重新執行,以控制這項威脅。
- Evaluation Awareness & Grader Gaming —— 訓練時期的威脅版本:能推理 grader 的模型,可以滿足成功的表象。解耦不會移除這項能力,但會拒絕讓 optimizer 直接取得 grader 的 feedback signal 來進行最佳化。
- Formal proof search —— 極端案例:Lean compiler 不只與 prover 解耦,還是可靠的 evaluator;這正是 proof-search 迴圈能以完全自主方式運作,而 agent 的 eval-fix 迴圈仍受 human-gated 的原因。
殘留的缺口#
將評分解耦後,仍有兩種耦合存在。第一是指標選擇:在 flywheel 示範中,之後提出修正方案的同一個 coding agent,也負責設計 custom rubric——optimizer 不能替自己的工作評分,但仍能決定要評分的內容。第二是lineage:如果獨立 evaluator 與受測 agent 來自同一個模型家族(Gemini 為 Gemini 建構的 agent 評分),那麼 judge-lineage bias 仍會穿過這道架構拆分而存在。解耦是必要條件,而非充分條件;它只是將信任問題提升一個層級,而不是消除問題(Loop Engineering 也指出同樣的退化:誰來驗證 verifier?)。
還有第三個更基本的缺口:獨立 evaluator 仍然必須是有效的。 解耦帶來的是獨立性,不是正確性——另一個獨立的 judge 可能完全可重現,卻仍系統性地錯誤。Norman et al. (2026) 以一致性–偏差悖論具體說明了這點:一個 test-retest 達 0.99 的 judge,仍可能帶有 0.19 的 position bias,確定性地偏好排在第一位的答案。這樣的 judge 能通過每一項「是否穩定/是否解耦?」檢查,卻仍會回傳無效裁決。因此,「optimizer 絕不替自己的工作評分」是第一項不變量;「grader 已完成 chance correction 並接受 bias audit」(Minimum Viable Validation Protocol)是第二項,而兩者互不蘊含。
相關連結#
- Agent Quality Flywheel —— 將這項規則表述為其 eval-fix 迴圈的設計不變量
- Reward Hacking —— 這項規則要防止的失敗模式,從訓練迴圈移至開發迴圈
- Loop Engineering —— maker/checker sub-agent 拆分與
/goal的獨立 stop-checker;同一規則在實務中的形式 - LLM-as-a-Judge —— 自我評分與 lineage bias 是問題在 judge 端的表述;獨立 judge 選擇則是 benchmark 端的緩解方式
- Evaluation Awareness & Grader Gaming —— grade-gaming 的模型內部版本;結構性解耦能抑制但無法消除它
- Verification as the New Bottleneck —— 解耦的 evaluation 讓 verification 足夠可信,因而得以委派
- LLM-Judge Validation —— 解耦所假設、卻不會自行提供的 validity layer:獨立 judge 可能可靠地出錯(一致性–偏差悖論),因此還必須完成 chance correction 與 bias audit
推導#
- Single General Agent vs. Multi-Agent Coding Architecture —— 這項規則正是多 agent coding 設定中的「testing/QA/reviewer」agent 之所以不可或缺的原因,因為它是獨立的:評估獨立性是一項結構性的(Goodhart)不變量,能在模型改進後持續存在,不像手工設計的任務結構那樣容易改變
開放問題#
- 解耦是否需要向上游延伸到指標設計?自行撰寫 rubric 的 optimizer,比只讀取分數的 optimizer 擁有更隱蔽的操弄管道。
- 多大的獨立性才足夠——不同模型家族、不同供應商,或不同的檢查模態(model judge、compiled test、production telemetry)?
資料來源#
- Driving the Agent Quality Flywheel from Your Coding Agent- Google Developers Blog —— 「The optimizer never grades its own work」段落(
vendor-claim) - Reliability without Validity: A Systematic, Large-Scale Evaluation of LLM-as-a-Judge Models Across Agreement, Consistency, and Bias —— Norman et al.(arXiv 2606.19544,2026 年 6 月,
empirical):一致性–偏差悖論(§4.7)——獨立且可重現的 judge 仍可能系統性偏誤;MVVP(§5.3)是解耦所遺漏的 validity check
Cited by 29
- Agent-Authored Harness Optimization×4
HarnessBank ablates a significance test off an evaluator that stays exogenous in every arm. Guo et…
- Oversight When the Signals Give Out: the Activation Fallback and the Taste Reward×4
The optimizer is already modeling the grader. Evaluation Awareness And Grader Gaming and the NLA…
- Recursive Self-Improvement×4
That cuts against the verification leg above in a specific way worth flagging. Guo et al. establish…
- Same-Model Review Blindness×4
Adversarial review inside one family. The Bun port ran implementer, two adversarial reviewers and a…
- Agent Harness Engineering×3
Optimizer Evaluator Decoupling — the rule the observer-independence result above obeys, and the one…
- Agent Quality Flywheel×3
Someone ran that direction anyway. Cline's July 2026 campaign (Agent Authored Harness Optimization,…
- Evaluation Awareness & Grader Gaming×3
self authored verification unreliable — Guo et al. (Chinese Academy of Sciences, arXiv 2607.24300,…
- Failures That Look Like Success×3
self authored verification unreliable — Guo et al. (Chinese Academy of Sciences, arXiv 2607.24300,…
- Single General Agent vs. Multi-Agent Coding Architecture×3
Optimizer Evaluator Decoupling: "the thing that proposes a change never grades that change." An…
- Cost-per-Task Over Cost-per-Token×2
Optimizer Evaluator Decoupling — the advisor strategy is that invariant reached from the cost side;…
- Cursor×2
Optimizer Evaluator Decoupling — Cursor's review lenses add the axis Bun's spec held fixed: what…
- Dynamic Workflows: An Algebra for Agents×2
The review() in the pseudocode is a specific, unusually well-specified design — the strongest…
- Gemini Enterprise Agent Platform×2
Optimizer Evaluator Decoupling — the evaluation service is the independent grader the rule requires
- Knowledge-Centric Self-Improvement×2
Optimizer Evaluator Decoupling — satisfied at the outcome layer (a deterministic external benchmark…
- Loop Engineering×2
Optimizer Evaluator Decoupling — the maker/checker split and /goal's separate stop-checker, stated…
- Open Questions Backlog×2
Optimizer Evaluator Decoupling: How much independence is enough — different model family, different…
- Parallel Agent Orchestration×2
Surface tension worth flagging. The same document says writer-verifier patterns are effective and…
- Reward Hacking×2
Optimizer Evaluator Decoupling — the structural countermeasure in eval-fix loops: deny the…
- Agent Review Comment Resolution
Optimizer Evaluator Decoupling — the deployed instance of the split at population scale: a reviewer…
- AI-Driven Formal Proof Search
Optimizer Evaluator Decoupling — the compiler is the limit case of that rule: an evaluator not…
- Deterministic Pre-Execution Gates
Optimizer Evaluator Decoupling — an exogenous verifier one layer down from where that page usually…
- LLM-as-a-Judge
Optimizer Evaluator Decoupling — the self-grading/lineage caveat elevated to an architectural rule:…
- LLM-Judge Validation
Optimizer Evaluator Decoupling — decoupling makes the evaluator independent but not valid; a…
- Agent Systems & Harness Engineering
Optimizer Evaluator Decoupling — The architectural rule in eval-fix loops that whatever proposes a…
- Reference-Free Judge Over-Crediting
Optimizer Evaluator Decoupling — the same rule with a different quantity decoupled, and the axis…
- Review as the Control Point
Optimizer Evaluator Decoupling — the architectural precondition for delegating review at all: the…
- Risk-Tiered Auto-Approval
Optimizer Evaluator Decoupling — the same source's reviewer-panel pattern applies the rule with an…
- Stopping Under a Noisy Verifier
Optimizer Evaluator Decoupling — the layer above. Decoupling gets you a verifier the optimizer did…
- Unproductive Self-Verification
Optimizer Evaluator Decoupling — self-verification's other failure mode, and it wants the opposite…
Related articles
- Verification as the New Bottleneck
Fiona Fung: coding is no longer the bottleneck — verification, review, maintenance are; shift-left; TDD loses its tax;…
- Open Questions Backlog
_456 actionable open questions across 205 pages · 107 predictions · 9 notes · 147 in progress · 69 watching (entities),…
- Agent-Authored Harness Optimization
An agent runs the whole eval-fix loop on its own harness — read traces, hypothesize, patch, re-run. Three instances dis…
- Dynamic Workflows: An Algebra for Agents
Claude Code's sandboxed orchestration primitive: Claude writes and runs a program that composes agents in sequence and…
- LLM-as-a-Judge
Using one LLM to grade another's outputs against criteria/rubrics; DRACO's protocol is per-criterion binary MET/UNMET +…
