資料來源#
摘要#
Fiona Fung 執行 Claude Code + Cowork 工程工作的核心主張是:多年來,工程頻寬一直是昂貴的資源——規劃、審查與流程都是為了保護它而存在。當代理式編碼讓編碼變得廉價後,瓶頸轉移到了驗證、審查與維護。「在 Claude Code 團隊,編碼真的已經不再是緩慢的部分了。」新的稀缺資源是「對變更正確無誤的信心」——而且正因為頻寬(以及因此而來的吞吐量)暴增,它變得更加稀缺。
為什麼驗證如今成了限制因素#
三股力量匯聚在一起:
- 數量。 頻寬大幅增加,以至於「我們必須更加注意:它是否正確」。
- 角色模糊化。 現在有更多人(設計師、經理、PM)會提交變更,因此每個人都需要確信自己的變更是正確的。
- 維護成本。 更高的吞吐量意味著需要維護的內容更多——維護成本成為首要考量,而不再是事後才想到的問題。
這是 Karpathy 的 The Verifiability Thesis(「LLM 會自動化你能驗證的事物」)在組織層級的映射,也是 Harness Shrinkage as Models Improve 的需求面(提示鷹架縮小;機械式驗證仍然承擔核心負荷)。
TDD 不再有額外負擔#
這種轉變的一個鮮明跡象是:TDD 過去讓人覺得像「吃花椰菜」——先寫出失敗的測試,確認它失敗,再修正。Fung 發現,使用 Claude 後「有趣、愉快得多……它消除了測試驅動開發的額外負擔」。經濟學已經翻轉:當寫測試幾乎不花成本時,支撐驗證的紀律(先有一個可證明會失敗、接著通過的測試)就只帶來好處。(參見 tdd/紅綠重構的紀律;先寫失敗測試的步驟就是驗證器。)
左移#
她反覆使用的詞是:左移——透過自動化,在更接近問題源頭的地方捕捉問題,而不是等客戶遇到問題後才處理。「有什麼比我先遇到 bug 更好?就是建立自動化機制,在更接近源頭的地方捕捉它。」隨著吞吐量上升,驗證要跟上進度的唯一方式,就是變得自動化且及早進行,而不是手動且延後處理。
誰來審查——以及人類參與其中的界線#
在推出 Claude Code 自己的程式碼審查功能前,「你們要如何跟上程式碼審查?」是她最常被問到的問題。答案是:Claude Code review 處理風格、lint、明顯的 bug,以及規格漂移(如果你把規格提交到程式碼庫中,「Claude 非常擅長針對規格漂移進行驗證」)。但在人類真正重要的地方,人類仍然參與其中:法律審查、風險容忍度、信任邊界——「信任但要驗證,以及在人類能帶來必要專業的地方讓人類參與」。分工方式是:把機械式驗證自動化,將人類判斷保留給風險與信任邊界的決策。(參見 Deep Modules for Agents:在全新脈絡中進行審查。)
衡量轉變(以及一個陷阱)#
她關注的訊號包括:新人上手時間 ↓、PR 週期時間 ↓、Claude 協助的提交 ↑(「我已經好幾個月沒看過不是 Claude 協助完成的提交了」)。陷阱在於:不要只看端到端的 PR 週期時間——要把它拆成漏斗中的各個區段。如果週期時間沒有下降,原因可能不是 AI 採用率低;也可能是CI/建置系統在新的吞吐量下塞車。而且吞吐量不是目標——「想辦法衡量你實際想解決的問題」,不要只衡量速度。
相關連結#
- Fiona Fung — 這項論點的作者
- The Verifiability Thesis — Karpathy 的「將你能驗證的事物自動化」是模型層級的原因;本文則是組織層級的後果
- Harness Shrinkage as Models Improve — 它所確認的綜合結論:鷹架會縮小,機械式驗證不會
- Evals as Product Spec — Cat Wu 的 evals 是編碼成產品規格的驗證;產品經理側的配套觀點
- Code as Source of Truth — 將規格檢查進儲存庫,才能讓 Claude 驗證規格漂移
- Building Is Cheap, Arguing Is Expensive — 上游的另一半:生成很便宜,因此成本集中在驗證(以及判斷)上
- Claude Code Auto Mode — 自動核准分類器是權限層的驗證自動化
- Deep Modules for Agents — 在全新脈絡中進行審查,是程式碼審查層級的驗證品質提升
- AI Brain Fry — 如果驗證仍然手動進行,風險就在於:隨著數量增加,監督疲勞會提高錯誤率
- AI-Driven Formal Proof Search — 極端案例:由編譯器擔任驗證器,因此瓶頸完全機械化
- Recursive Self-Improvement — 組織的 Amdahl's law:隨著生成加速,人類程式碼審查成為 Anthropic 的新瓶頸——這是 AI 建造 AI 規模下的同一論點
- AI Accelerating AI Development — 相互佐證的資料:自動化的 Claude 審查器,原本可以在合併前抓到過去生產事故背後約 1/3 的 bug
- Research Taste as the Human Bottleneck — 當人類無法像 Claude 生成內容那樣快速地審查/判斷時,判斷就成為關鍵限制因素——這是驗證的更高層次形式
- Loop Engineering — 製作者/檢查者子代理分工是其五個基元之一,而
/goal的獨立模型停止檢查,正是將本文論點應用於「完成」決策;Osmani 的「你的工作是交付你已確認能運作的程式碼」,以及無人值守迴圈的審查頻寬上限,都是迴圈層級的同一限制 - Acceleration Whiplash — Faros AI 的業界遙測資料佐證了這個瓶頸(PR 審查中的中位時間 +441.5%,31.3% 的 PR 在沒有審查的情況下合併),但也細化了修正方向:透過提升撰寫品質來紓解瓶頸,而不是擴大審查層
- Review as the Control Point — 本論點底下的機制圖:審查是決定編碼代理如何影響軟體的地方,但影響的正負號由團隊決定(審查者專業、處置方式、流程),而不是由 AI 決定——其 P8/P9/P17 為本頁開放問題所詢問的速度/安全性提供了答案
- AI as Primary Author — 當 AI 撰寫大多數程式碼後,品質差距源自審查上游——「這是撰寫問題,而不是審查問題」
- Conversation Artifacts — 對對話產生的輸出進行分類,是將人類必須審查的內容加以儀表化的一步;產出物就是可審查的單位
- The Three Loops of AI-Native Building — 直接的異議。 Andrew Ng 報告了相反的動向:自我測試代理降低了開發者的 QA 負擔(「我們需要花在這項功能上的時間大幅減少」),讓人類提升到產品決策,而不是困在審查中。兩者都是
practitioner-opinion;範圍不同(從零到一的個人建置 vs 生產組織),而 Faros 的遙測資料 是證據力更高的決勝依據,在組織案例上站在 Ng 的對立面 - Unknowns as the Agentic Bottleneck — 上游一步的瓶頸:驗證會問「這是對的嗎?」,未知會問「我是否曾經說明什麼才是對的?」;測驗閘門將製作者/檢查者分工應用到人類身上
推導#
- When Does Verification Quality Determine Whether AI Automation Works? — 將此瓶頸一般化為驗證品質階梯:Lean/形式化證明、軟體 CI、漏洞重現,以及嘈雜的判斷任務
開放問題#
- Fung 自己提出的開放問題:「你會把全自動審查推進到多遠?」——速度/安全性的平衡點在哪裡?如何讓人類保持信心,同時不重新引入審查瓶頸?Review as the Control Point 進一步銳化了這個問題:完整自動化可靠地提高審查吞吐量並降低延遲(其 P8),但它對程式碼品質與安全性的影響仍有爭議(P9),而審查治理政策對延遲的影響,會因校準方式而反轉——只限制重大變更的風險分級政策會降低延遲,一刀切政策則會提高延遲(P17)。因此,「要推進到多遠」沒有單一答案:安全前沿由自動化審查器的能力與流程設計決定(該頁三個調節因素中的兩個),而不是由固定的旋鈕決定。
- 如果 CI/建置是隱藏的塞車點,驗證基礎設施(測試執行器、CI 容量)是否會成為 AI 原生組織真正的資本支出?
資料來源#
- Running an AI-native engineering org
- Thread by @AndrewYNg — Andrew Ng,《The Batch》(2026-06-30),
practitioner-opinion:這項異議指出,在從零到一的建置中,自我測試代理降低了人類的驗證負擔,而不是提高它
Cited by 88
- Human-in-the-Loop Boundaries×5
Verification As The New Bottleneck says correctness confidence is now the bottleneck, so mechanical…
- Loop Engineering×4
Verification is still yours. "A loop running unattended is also a loop making mistakes unattended."…
- The PRD-Replacement Spectrum at AI-Native Speed×4
The deep precondition behind the whole right half is Verification As The New Bottleneck: "generate…
- When Does Verification Quality Determine Whether AI Automation Works?×4
The failure mode is not "AI cannot code." The failure mode is that code volume outruns verification…
- Acceleration Whiplash×3
This is a productive refinement of Verification As The New Bottleneck: Faros agrees verification is…
- Organizational Complements to AI×3
The evaluation queue lengthened 2.8×. Median interview→offer-decision time went from 2.62 days…
- Research Taste as the Human Bottleneck×3
Even if taste stays human, it becomes the binding constraint — the Amdahl's-law bottleneck of the…
- Addy Osmani×2
The code agent orchestra / adversarial code review — the maker/checker split that Verification As…
- Agent-Authored Harness Optimization×2
Verification As The New Bottleneck — the human PR review is the last decoupled check in this loop…
- Agent-Generated Test Quality×2
Verification As The New Bottleneck — flaky agent tests are the mechanism behind Fung's warning that…
- Agentic Honesty & Diligence×2
These are exactly the failure modes that make autonomous agentic coding risky: when a model writes…
- AI Accelerating AI Development×2
Verification As The New Bottleneck — the automated reviewer and "review became the new bottleneck"…
- AI as Primary Author×2
Verification As The New Bottleneck — humans-as-reviewers-not-creators is exactly the bottleneck…
- Building Is Cheap, Arguing Is Expensive×2
The thing the team reduced: "really in-depth planning and design docs. Most of our discussions…
- Code as Source of Truth×2
Verification As The New Bottleneck — the cause: high throughput stales docs; spec-in-repo enables…
- Confident But Unsure×2
Verification As The New Bottleneck — analysis-shaped wrong answers are the most expensive kind to…
- Controlled Variance: AI's Edge as Reduced Dispersion×2
Automating the interview made scheduling nearly instant (the agent is available 24/7) and made the…
- Cost-per-Task Over Cost-per-Token×2
So the practical rule when reading any cost-per-task claim about an agent campaign: ask which cost…
- Deep Research Agents×2
Verification As The New Bottleneck — factual accuracy / citation are the weakest axes across all…
- Dogfooding as Product Discipline×2
Once coding is cheap (Verification As The New Bottleneck), the constraint shifts to knowing what's…
- Dynamic Workflows: An Algebra for Agents×2
The pattern to carry: the escaped defects were the ones where the test oracle and the build…
- Failures That Look Like Success×2
A third incident in the same catalogue is the quiet version, and the most costly: an agent hit an…
- Fiona Fung×2
Bottlenecks moved — from engineering bandwidth to verification, review, and maintenance…
- Is Human Review of AI-Authored Code Still a Real Control, or Already Rubber-Stamping?×2
Fung's question assumes a single automation frontier to push. The evidence decomposes it…
- Implementation Abundance Inverts Product Work×2
Andrew Ambrosino's (OpenAI Codex) framing of what agentic coding does to product process: when…
- Open-Ended Discovery Harnesses×2
Verification As The New Bottleneck — the case study's self-reported risk: a harness that multiplies…
- Open Questions Backlog×2
Verification As The New Bottleneck: If CI/build is the hidden jam, does verification infrastructure…
- Optimizer–Evaluator Decoupling×2
The generalizable note: when the artifact under optimization is the eval substrate, the split has…
- Parallel Agent Orchestration×2
The paper grounds parallelism in the same property that makes coding the leading edge of agentic…
- Planning / Execution Division of Labor×2
Anthropic's 400K-session study supplies the empirical shape of human–agent collaboration in agentic…
- Product Velocity as Moat×2
Velocity has always helped startups; what makes it a moat now is the AI-native cost structure. When…
- Recursive Self-Improvement×2
A recurring brake across futures 2–3: speeding up one part of a process just shifts the bottleneck…
- Single General Agent vs. Multi-Agent Coding Architecture×2
Verifier availability sets the ceiling. Perfect verifier (Lean, a passing test suite) → high…
- The Three Loops of AI-Native Building×2
The human didn't get removed from the loop; they got promoted out of QA. Notice this cuts against…
- Unknowns as the Agentic Bottleneck×2
Verification As The New Bottleneck — the bottleneck one step upstream: verification asks "is this…
- Unsanctioned Action in Capability Evaluations×2
The chained version is worse. When the agent briefly had code execution inside the bystander's…
- Agent Quality Flywheel
Verification As The New Bottleneck — the bottleneck this tooling attacks: grading, failure…
- Agent Review Comment Resolution
Verification As The New Bottleneck — an adoption datum for "how far do you push fully automated…
- Agentic Prompt Injection
It is agent data injection, not instruction injection — the forged artifact is a trusted status…
- AI Brain Fry
Verification As The New Bottleneck — the review/verification burden is where oversight fatigue…
- AI-Driven Formal Proof Search
Verification As The New Bottleneck — compiler-verified proofs are the purest case of verification…
- AI-Native Product Org Bottlenecks
Verification As The New Bottleneck — broader engineering-side version of the same shift from…
- Andrew Ng
QA was the job that went away. "Last year, a lot of developers (including me) were acting as the QA…
- Anthropic
Fiona Fung — leads engineering + product for Claude Code + Cowork; author of the…
- Authority and Audit Survive Abundance
Self-reported attribution is model output. A model asked which span of a stuffed window grounded…
- Automated Failure Attribution
Verification As The New Bottleneck — the bottleneck extended past accept/reject. Deciding a run…
- Bun
Verification As The New Bottleneck — Bun's language-independent test suite is the substrate that…
- Cat Wu
Verification As The New Bottleneck — her "ten great evals" / "push to 100%" stances are the…
- Claude Code
Verification As The New Bottleneck — Fiona Fung: on the Claude Code team coding is no longer the…
- Claude Code Auto Mode
Verification As The New Bottleneck — auto-mode's classifier shifts the verification burden to…
- Configurable Human Participation
Verification As The New Bottleneck — the Feedback channel (evaluate/correct intermediate output)…
- Conversation Artifacts
Verification As The New Bottleneck — a legible artifact is what a human must review; classifying…
- Conversation-to-Delegation Shift
Verification As The New Bottleneck — as use shifts from asking to delegating, the human role moves…
- Crystallizing Agent Work into Workflows
Verification As The New Bottleneck — the auto-generated acceptance tests are what make promotion…
- Deep Modules for Agents
Verification As The New Bottleneck — reviewer-in-fresh-context at the module interface is a…
- Deterministic Pre-Execution Gates
Verification As The New Bottleneck — the cheapest possible tier of verification, placed before the…
- Document Parsing as the Retrieval Bottleneck
Verification As The New Bottleneck — why bbox grounding is the interesting ParseBench dimension: an…
- Documented Agent Incidents (METR Catalogue)
Verification As The New Bottleneck — INC-038 is the bottleneck failing quietly: the human ran a…
- DRACO Benchmark
Verification As The New Bottleneck — factual-accuracy weakness across all systems is verification…
- Efficiency Debt of AI-Generated Code
Verification As The New Bottleneck — the class of defect the bottleneck cannot absorb: not a matter…
- Evals as Product Spec
Verification As The New Bottleneck — Fiona Fung's org-level claim that verification (which evals…
- Faros AI
Verification As The New Bottleneck — its findings supply external telemetry for Fiona Fung's…
- Harness Build-vs-Buy
Verification As The New Bottleneck — the 866-bug-fixes-a-year figure is the maintenance half of…
- Harness-Induced Belief Divergence
Verification As The New Bottleneck — verification treated as an evidence channel rather than a…
- Harness Shrinkage as Models Improve
Verification As The New Bottleneck — Fiona Fung's org-level corollary: as the generation harness…
- Layerwise Omission Attribution
Verification As The New Bottleneck — omission as the hardest case for verification: there is no…
- LLM-as-a-Judge
Verification As The New Bottleneck — LLM-as-a-judge is one (imperfect) answer to the…
- LLM-Judge Validation
Verification As The New Bottleneck — LLM-judge validation is the quality-control layer under one…
- AI Coding Practice
Verification As The New Bottleneck (hub) — Fiona Fung: coding is no longer the bottleneck —…
- Output Length Calibration
Verification As The New Bottleneck — longer deliverables are paid for by the reviewer; uncalibrated…
- Post-Acceptance Edit Behavior
Verification As The New Bottleneck — the bottleneck's cheapest stage, and the shortest feedback…
- Repository Exploration Subagent
Verification As The New Bottleneck — localization/exploration is the upstream sibling of…
- Returns to Expertise in Agentic Coding
Verification As The New Bottleneck — "what they ask Claude to verify" is one of the three expertise…
- Review as the Control Point
Verification As The New Bottleneck — the control-point claim is the bottleneck thesis with a…
- Risk-Tiered Auto-Approval
Verification As The New Bottleneck — a partial, deployed answer to that page's "how far do you push…
- RSI Growth Curves: Which Friction Binds First?
1. Already binding (organizational scale, mid-2026). The Amdahl's-law / verification-and-oversight…
- Same-Model Review Blindness
Verification As The New Bottleneck — "how far do you push automated review" gains an axis…
- Security Debt of Agent-Generated Code
Verification As The New Bottleneck — the measured floor under the bottleneck: on the one smell…
- The Solo-Authorship Rebound
Verification As The New Bottleneck — the solo author is the sole verifier of work an LLM executed,…
- Stopping Under a Noisy Verifier
Verification As The New Bottleneck — the bottleneck given a coefficient: verification is not just…
- Systems Thinking Over Specialization
Verification As The New Bottleneck — humans "reason and rationalize" over agent-scale output; the…
- Telemetry vs. Survey Measurement
Verification As The New Bottleneck — Fiona Fung's warning to break PR-cycle-time into funnel chunks…
- The Tragedy of the Cognitive Commons
Verification As The New Bottleneck — the Validation Tether is the bottleneck's precondition:…
- Unproductive Self-Verification
Verification As The New Bottleneck — the inversion: verification is supposed to be the human's…
- The Verifiability Thesis
Verification As The New Bottleneck — Fiona Fung: once coding is cheap, verification (not…
- Verifying Without a Compiler: Cowork's Harness vs Claude Code's, and Why the Slice Verifier Stays
Claude Code's harness leans on a post-hoc deterministic verifier stack. Tests, compilers, linters,…
- Vibe Coding vs. Agentic Engineering
Verification As The New Bottleneck — Fiona Fung's org-level account of "preserve the quality bar…
- Why AI Lags at Design
Verification As The New Bottleneck — the general shape: capability races ahead where verification…
Related articles
- Harness Shrinkage as Models Improve
Prompt scaffolding shrinks each model release; Cat Wu's pruning discipline; Boris Cherny "100 lines of code a year from…
- Claude Code
Anthropic's agentic coding product; created by Boris Cherny late 2024; TypeScript/React on Bun (itself Claude-rewritten…
- Agentic Technical Debt
Debt that *compounds* (not just accumulates) because each agentic-coding session re-derives architectural decisions wit…
- Boris Cherny
Creator of Claude Code at Anthropic; phone-driven workflow with hundreds of agents; primary advocate of `/loop` primiti…
- Open Questions Backlog
_456 actionable open questions across 205 pages · 107 predictions · 9 notes · 147 in progress · 69 watching (entities),…
