AI 編碼智能體令提交量增加 240%,軟件發布量僅升 30%
AI Coding Agents Raised Commits 240%. Releases Rose Only 30%
一項涵蓋超過 500,000 名 GitHub 開發者的分析估算,AI 編碼智能體令提交量增加 240%,但項目層面的增幅為 80%,軟件發布量只增加 30%。文章綜合多項研究指出,生成量未必轉化為交付,審核、測試和返工等驗證工作會成為瓶頸。文中建議以通過審查、測試、安全和部署門檻的變更衡量產出,並追蹤驗證時間相對於生成時間的比例。
Autonomous coding agents lifted commit volume 240% across more than 500,000 GitHub developers, yet software releases rose only 30% (Source: National Bureau of Economic Research, 2026). The Philippines is buying this capability fast: 20.7% of working-age Filipinos used a generative AI product in the second quarter of 2026, above the 18.8% global average (Source: Microsoft AI Economy Institute, 2026). The Department of Science and Technology launched its National Artificial Intelligence Center for Research and Innovation in February 2026 with a P2.6 billion budget (Source: Rappler, 2026). Generation is no longer scarce, and the design problem moves downstream.
The Attenuation Curve Is the Architecture Finding
The working paper used a matched event study combining GitHub data with AI usage telemetry to estimate the 240% commit lift for autonomous agents (Source: National Bureau of Economic Research, 2026).
The gains shrink down the production hierarchy: the 240% commit effect falls to 80% for projects and 30% for releases (Source: National Bureau of Economic Research, 2026). The authors call this the weak-link hypothesis, with an estimated elasticity of substitution of 0.23 between AI and human effort (Source: National Bureau of Economic Research, 2026).
Across four major software marketplaces, the same data shows a sharp increase in new apps and no increase in total usage (Source: National Bureau of Economic Research, 2026). Applications got easier to produce while demand did not.
Controlled Trials Split by Setting
Three field experiments at Microsoft, Accenture, and an anonymous Fortune 100 company, pooled across 4,867 developers, found a 26.08% increase in completed tasks with an AI coding assistant (standard error 10.3%), and less experienced developers showed higher adoption and greater gains (Source: Management Science, 2026).
METR ran a randomized controlled trial with experienced open-source developers on issues in mature repositories. Developers expected AI to make them 24% faster, and they took 19% longer. The perception gap never closed: afterward they still believed AI had sped them up (Source: METR, 2025).
METR's February 2026 follow-up covered 57 developers and more than 800 tasks, estimating speedup at -18% for returning participants and -4% for newcomers, both with confidence intervals crossing zero. The researchers called that weak evidence, because many developers avoided tasks they did not want to do without AI (Source: METR, 2026).
Broad telemetry is flatter. A classifier applied to more than 30 million GitHub commits estimated AI writes 29% of Python functions in the US, quarterly code contributions up 3.6%, and that experienced programmers captured nearly all of those gains (Source: Science, 2026).
The Verification Tax Reaches the Budget
A March 2026 DORA analysis of 1,110 Google software engineer responses found that time saved in code generation is frequently re-allocated to auditing and verification, and links higher AI adoption to both higher throughput and higher instability (Source: DORA, 2026). The 2025 DORA report found 90% of technology professionals use AI at work, over 80% believe it increased their productivity, and 30% report little to no trust in AI-generated code (Source: DORA, 2025). Adoption and distrust describe the same population.
A Monte Carlo simulation of 432 mixed human-and-agent team scenarios found escaped bugs the dominant cost driver, with 61% of the agent budget going to finding bugs (Source: Herbold, 2026). Agent capability mattered more than token price, because better outputs cut downstream quality assurance costs. Teams with high-capability agents were typically cheaper than human-only teams, while low-capability agents offered no consistent advantage, and at realistic velocities the optimal teams still spent more on developers (Source: Herbold, 2026).
Model selection is a lever; human engineering capacity is the larger line item.
The Philippine Constraint Is Not Tooling
Microsoft's Q2 2026 diffusion data puts the Philippines at 20.7% of the working-age population using generative AI, up from 20.1% in the first quarter, above the 18.8% global average but behind Singapore at 64.3%, Vietnam at 27.9%, and Malaysia at 22.6% (Source: Microsoft AI Economy Institute, 2026). The Philippine Daily Inquirer reported the country at 56th of 147 economies, and at 17.1% in the first half of 2025 (Source: Philippine Daily Inquirer, 2026).
Corporate readiness lags individual adoption. BCG's Philippine AI Maturity Report, drawn from 35 Philippine companies, found 60% of surveyed organizations still in early stages of adoption, the country nearly two years behind the global maturity curve, and Philippine labor cost running 70% below the Asia-Pacific average and 80% below the global average (Source: Boston Consulting Group, 2026).
That is why the attenuation curve matters here. The Philippines competes on cost, and many AI-native players emerged only after 2022 and already run at significant scale with small teams (Source: Boston Consulting Group, 2026). If output cannot ship, the cost advantage gets consumed by review and rework.
Architecture Changes That Match the Bottleneck
Four moves follow from the evidence for teams shipping agents to production.
- Count production-qualified changes, not commits. Credit a change only after it clears review, test, security, and deployment gates for its risk class. (Source: Bhati, 2026)
- Move AI feedback to the author, not the reviewer. Feedback given during writing is cheaper to act on than a reviewer receives later. (Source: DORA, 2026)
- Force small batches. Large AI-generated changes reviewed as one unit push verification cost back into the queue. (Source: DORA, 2026)
- Instrument the verification ratio. Track review and rework minutes against generation minutes, and treat that ratio as the tax.
FAQ
Q: Does the 240% figure mean coding agents are not worth using?
A: No. Commit volume is a poor proxy for delivered software. The same analysis found releases up 30% and no increase in app usage across four marketplaces, which locates the constraint rather than dismissing the tool. (Source: National Bureau of Economic Research, 2026)
Q: Is the 19% slowdown still the current finding?
A: Not as a headline. METR's February 2026 update flags the early-2025 result as out of date, and its follow-up could not reliably estimate a speedup because many developers avoided tasks they did not want to do without AI. (Source: METR, 2026)
Key Takeaway
Generation is cheap, qualification is not, and architecture now owns the qualification step. The teams that win this cycle stopped counting commits and started counting changes that survived the gates. Pick one metric your team reports this quarter and replace it: will your next review ask how much code was written, or how much shipped?
Sources
- Writing Code vs. Shipping Code - National Bureau of Economic Research
- Early-2025 AI and Developer Productivity - METR
- Changing our Experiment Design - METR
- Effects of Generative AI on High-Skilled Work - Management Science - SSRC Just Tech
- Who is using AI to code? - Science - arXiv preprint
- State of AI-assisted Software Development - DORA
- Balancing AI tensions - DORA
- What Should a Software Team Spend on Agents? - Herbold, arXiv
- Beyond Code Generation - Bhati, arXiv
- Global AI Diffusion Q2 2026 - Microsoft AI Economy Institute
- AI use rises in PH amid Asean push - Philippine Daily Inquirer
- The Philippine AI Maturity Report - Boston Consulting Group
- What is NAICRI? - Rappler
來源:DEV Community · dev.to
