Every source behind the report, with its locator, method, population, quality tier, commercial interest, key limitation and, where it could be captured, the verbatim sentence that supports the claim.
26 of 26 sources carry the verbatim sentence that supports the claim. 12 of 26 are authored or funded by an organisation with a commercial interest in the finding.
State of AI-assisted Software Development 2025
90% of respondents use AI at work; >80% report perceived productivity improvement; 30% report little/no trust in AI-generated code; 59% observe positive impact on code quality.
| Population | Nearly 5,000 technology professionals plus qualitative research |
|---|---|
| Method | Global survey + qualitative research |
| Key limitation | Self-reported outcomes and cross-sectional organizational evidence; causal interpretation should be limited. |
| Commercial interest | Google Cloud sells the AI development tooling whose adoption this report measures. |
| Licence | CC BY-NC-SA 4.0 |
| Locator | https://services.google.com/fh/files/misc/2025_state_of_ai_assisted_software_development.pdf · archived |
| Supporting sentence | “"The majority of survey respondents (90%) use AI as part of their work and believe (more than 80%) it has increased their productivity. Yet a notable portion (30%) currently report little to no trust in the code generated by AI, indicating a need for critical validation skills." And: "a majority (59%) of survey respondents also observe that AI has positively impacted their code quality."” |
Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity
16 developers, 246 tasks; AI allowed condition took 19% longer, while developers expected 24% faster and perceived ~20% faster.
| Population | Experienced open-source developers; mature repositories; average ~5 years prior repo experience |
|---|---|
| Method | Randomized controlled trial |
| Key limitation | Narrow setting and early-2025 frontier tools; METR explicitly warns against broad generalization. |
| Licence | All rights reserved |
| Locator | https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ · archived |
| Supporting sentence | “Surprisingly, we find that when developers use AI tools, they take 19% longer than without—AI makes them slower.” |
Effects of Generative AI on High-Skilled Work: Three Field Experiments
4,867 developers; preferred estimate +26.08% completed tasks (SE 10.3%).
| Population | Developers at Microsoft, Accenture, and an anonymous Fortune 100 company |
|---|---|
| Method | Three randomized field experiments |
| Key limitation | Experiments began in 2022-era assistant paradigm; effect sizes may not transfer directly to newer agents or all work types. The +13.55% commits and +38.38% compilations figures could not be located at the recorded URL and have been removed pending a full-text locator. |
| Commercial interest | Run with Microsoft Research; Microsoft sells GitHub Copilot, the tool under study. |
| Licence | Proprietary - no licence declared |
| Locator | https://www.microsoft.com/en-us/research/publication/the-effects-of-generative-ai-on-high-skilled-work-evidence-from-three-field-experiments-with-software-developers/ · archived |
| Supporting sentence | “"when data is combined across three experiments and 4,867 developers, our analysis reveals a 26.08% increase (SE: 10.3%) in completed tasks among developers using the AI tool"” |
How much does AI impact development speed? Enterprise RCT
96 full-time Google software engineers; best estimate ~21% reduction in task time; confidence interval was large.
| Population | Google software engineers |
|---|---|
| Method | Randomized controlled trial |
| Key limitation | Single enterprise task and internal tooling; authors caution against extrapolating to ecosystem-wide productivity. |
| Commercial interest | Google Research studied Google engineers using Google tooling. |
| Licence | CC BY 4.0 |
| Locator | https://doi.org/10.48550/arXiv.2410.12944 · archived |
| Supporting sentence | “Our randomized controlled trial with 96 full-time Google software engineers contributes to this literature by sharing an estimate of the impact of three AI features on the time developers spent on a complex, enterprise-grade task.” |
Responsible AI pair programming with GitHub Copilot
95 developers; 45 with Copilot; Copilot group completed a JavaScript task ~55% faster.
| Population | Developers familiar with JavaScript |
|---|---|
| Method | Controlled experiment (vendor research) |
| Key limitation | Single constrained task; vendor-run study; not representative of end-to-end software delivery. Published 2023 on an experiment run with 2022-era Copilot: by this report's own tool-generation argument, the oldest tool generation in the evidence base and the least transferable to agentic tooling. |
| Commercial interest | GitHub studied its own product, Copilot. |
| Licence | Proprietary - no licence declared |
| Locator | https://github.blog/ai-and-ml/github-copilot/responsible-ai-pair-programming-with-github-copilot/ · archived |
| Supporting sentence | “These 45 developers completed the task 55% faster than the group without GitHub Copilot.” |
Does GitHub Copilot improve code quality?
202 valid submissions; Copilot users 53.2% more likely to pass all 10 unit tests; readability +3.62%, reliability +2.94%, maintainability +2.47%, conciseness +4.16%.
| Population | Study participants producing code reviewed against tests and quality dimensions |
|---|---|
| Method | Controlled study (vendor research) |
| Key limitation | Vendor-run study and constrained task; quality metrics do not prove production reliability. |
| Commercial interest | GitHub studied its own product, Copilot. |
| Licence | Proprietary - no licence declared |
| Locator | https://github.blog/news-insights/research/does-github-copilot-improve-code-quality-heres-what-the-data-says/ · archived |
| Supporting sentence | “Key findings from this study: Increased functionality: developers with GitHub Copilot access had a 53.2% greater likelihood of passing all 10 unit tests in the study, indicating that GitHub Copilot helps developers write more functional code by a wide margin.” |
2025 Developer Survey - AI
46% distrust AI accuracy vs 33% trust; 52% do not use agents or use simpler AI; among agent users ~70% report task-time reduction, 69% productivity increase, 17% collaboration improvement; 87% accuracy concern, 81% security/privacy concern.
| Population | Global Stack Overflow survey respondents; per-question sample sizes vary |
|---|---|
| Method | Developer survey |
| Key limitation | Self-selection and self-reported effects; agent questions have smaller respondent subsets. |
| Licence | Site (c) 2025 Stack Exchange Inc.; user contributions CC BY-SA; survey data ODbL |
| Locator | https://survey.stackoverflow.co/2025/ai · archived |
| Supporting sentence | “2 Developer tools Accuracy of AI tools More developers actively distrust the accuracy of AI tools (46%) than trust it (33%), and only a fraction (3%) report "highly trusting" the output.” |
Measuring AI Ability to Complete Long Tasks
50%-success task time horizon doubling approximately every 7 months over the preceding 6 years.
| Population | Self-contained software engineering, ML/research and cybersecurity tasks |
|---|---|
| Method | Benchmark + human baseline timing + logistic modeling |
| Key limitation | Tasks are cleaner and lower-context than much real-world engineering; extrapolation to 2030 is uncertain. |
| Licence | All rights reserved |
| Locator | https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/ · archived |
| Supporting sentence | “"The length of tasks (measured by how long they take human professionals) that generalist frontier model agents can complete autonomously with 50% reliability has been doubling approximately every 7 months for the last 6 years."” |
Task-Completion Time Horizons of Frontier AI Models
Task-horizon measurements above ~16 hours are flagged as unreliable with the current task suite.
| Population | Frontier public language-model agents |
|---|---|
| Method | Time Horizon 1.1 benchmark methodology |
| Key limitation | LIVING PAGE: figures change without notice. Must be cited as “figures as of 20 August 2026” and re-checked quarterly. The seven-month doubling claim belongs to S10, not to this page. |
| Licence | All rights reserved |
| Locator | https://metr.org/time-horizons/ · archived |
| Supporting sentence | “(living page; no fixed prose claim quoted - see S10 for the doubling rate)” |
Running Codex safely at OpenAI
Operational design pattern: routine low-risk actions can proceed; higher-risk actions are made explicit and bounded.
| Population | Internal coding-agent deployment |
|---|---|
| Method | Operational practice / case study |
| Key limitation | Vendor-specific implementation; principles require adaptation to enterprise context. |
| Commercial interest | OpenAI reporting on its own internal use of its own models. |
| Licence | Proprietary - no licence declared |
| Locator | https://openai.com/index/running-codex-safely/ · archived |
| Supporting sentence | “Sandboxing and approvals Approvals and sandboxing work together.” |
How far can we push AI autonomy in code generation?
Qualitative experimental findings across agentic workflows for a Spring Boot application.
| Population | Thoughtworks experimental workflow |
|---|---|
| Method | Engineering experiment / practitioner research |
| Key limitation | Not a controlled population study; results depend on models, prompts and stacks available at the time. |
| Licence | All rights reserved |
| Locator | https://martinfowler.com/articles/pushing-ai-autonomy.html · archived |
| Supporting sentence | “sking versus TODO lists: What about parallelisation? How far can we push AI autonomy in code generation? We ran a series of experiments to explore how far Generative AI can currently be pushed toward autonomously developing high-quality, up-to-date software without human interven” |
SP 800-218A: Secure Software Development Practices for Generative AI and Dual-Use Foundation Models
Final NIST SSDF Community Profile for generative AI and dual-use foundation models.
| Population | Organizations developing AI/model-enabled systems |
|---|---|
| Method | Standards guidance |
| Key limitation | Guidance is intentionally high-level and requires organization-specific controls. |
| Licence | U.S. Government work - public domain in the United States |
| Locator | https://csrc.nist.gov/pubs/sp/800/218/a/final · archived |
| Supporting sentence | “SP 800-218A, Secure Software Development Practices for Generative AI and Dual-Use Foundation Models: An SSDF Community Profile | CSRC You are viewing this page in an unauthorized frame window.” |
OWASP Top 10 for Agentic Applications 2026
Globally peer-reviewed framework developed with 100+ experts/researchers/practitioners.
| Population | Agentic AI applications |
|---|---|
| Method | Community security framework |
| Key limitation | Threat framework, not incidence-rate measurement. |
| Licence | CC BY-SA 4.0 |
| Locator | https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/ · archived |
| Supporting sentence | “Developed through extensive collaboration with more than 100 industry experts, researchers, and practitioners, the list provides practical, actionable guidance to help organizations secure AI agents that plan, act, and make decisions across complex workflows.” |
Future of Jobs Report 2025
Over 1,000 employers representing 14M+ workers, 22 industry clusters, 55 economies. 39% of workers' existing skill sets expected to be transformed or become outdated over 2025-2030 (down from 44% in 2023, 57% in 2020). Software and Applications Developers ranks 4th of 15 fastest-growing jobs by projected percentage growth (Figures 2.2 and 2.3).
| Population | Global employer survey |
|---|---|
| Method | Global employer survey + labor-market modeling |
| Key limitation | Employer self-report of intentions, not measurement of outcomes. Rankings are percentage growth as a share of current employment in each role, not absolute headcount, so a small occupation can outrank a large one. Survey base is over 1,000 employers, which is small relative to the 14 million workers they are reported as representing. |
| Licence | All rights reserved ((c) 2025 World Economic Forum) |
| Locator | https://www.weforum.org/publications/the-future-of-jobs-report-2025/ · archived |
| Supporting sentence | ““On average, workers can expect that two-fifths (39%) of their existing skill sets will be transformed or become outdated over the 2025-2030 period.” // Figure 2.2, p.18, ‘Top fastest growing jobs’, in rank order: Big Data Specialists, FinTech Engineers, AI and Machine Learning Specialists, Software and Applications Developers, Security Management Specialists. Body text, p.18: “the fastest-growing job roles by 2030, in percentage terms … Leading the fastest growing jobs list are roles such as Big Data Specialist, FinTech Engineers, AI and Machine Learning Specialists and Software and Applications Developers.” // Preface, p.4: “over 1,000 employers – representing more than 14 million workers across 22 industry clusters and 55 economies”.” |
Industry and Occupational Employment Projections, 2024-34
Software developers (SOC 15-1252): 1,693.8k (2024) to 1,961.4k (2034), +267.7k / +15.8%. Ranked 24th of 832 occupations by projected percent change.
| Population | U.S. labor market |
|---|---|
| Method | Official employment projections |
| Key limitation | U.S.-specific and occupational projections cannot capture all role redesign within software teams. BLS blocks automated access (403). Figures require a human to open the Occupational Outlook Handbook and confirm 1.6938M -> 1.9614M and +15.8%. |
| Licence | Public domain - U.S. Government work |
| Locator | https://www.bls.gov/opub/mlr/2026/article/industry-and-occupational-employment-projections-overview.htm · archived |
| Supporting sentence | “Table 5, row 15-1252 Software developers: Employment 2024 = 1,693.8; Projected employment 2034 = 1,961.4; Projected numeric change 2024-34 = 267.7; Projected percent change 2024-34 = 15.8; Rank 24th of 832. (Figures in thousands.)” |
Spec-Driven Development and Agentic SDD
Official Spec Kit workflow: constitution -> specify -> clarify -> plan -> checklist -> tasks -> analyze -> implement -> converge; the core default is Spec -> Plan -> Tasks -> Implement.
| Population | AI coding-agent workflows using GitHub Spec Kit |
|---|---|
| Method | Official open-source methodology and product documentation |
| Key limitation | Methodology/tooling evidence rather than independent productivity proof; adoption and outcomes vary by project and organization. |
| Commercial interest | GitHub publishing on its own specification tooling. |
| Licence | MIT ((c) GitHub, Inc.) |
| Locator | https://github.com/github/spec-kit/blob/main/docs/reference/agentic-sdd.md · archived |
| Supporting sentence | “specify is strictly required before /speckit.” |
Developer Ecosystem Survey 2026 - AI coding agent adoption
90% use AI coding agents at work at least weekly; 68% daily. Claude Code 39% global / 47% US (18% in Jan 2026), primary tool for 31%; GitHub Copilot 21% (29% YoY prior); Codex 16% (3% in Jan 2026); Cursor 12% (18% in Jan 2026); JetBrains AI ~9%; OpenCode 7%; Google Antigravity 6%.
| Population | 15,000+ professional developers worldwide, ~90% developer/engineer titles |
|---|---|
| Method | Global survey, regional quotas, reweighted by region, employment status, language and product familiarity; fielded May-July 2026 |
| Key limitation | Vendor-run survey (JetBrains sells developer tools and one of the tools measured). Self-reported tool usage. No margin of error published. Re-tiered to B: this is a vendor-run survey, and the report’s own tier definition places vendor empirical work at B regardless of sample size. |
| Commercial interest | JetBrains sells developer tooling, including an AI assistant that appears in its own results. |
| Licence | Proprietary - no licence declared |
| Locator | https://blog.jetbrains.com/research/2026/08/ai-coding-agent-adoption-2026/ · archived |
| Supporting sentence | “90% of professional developers were using AI coding agents at work at least weekly in one form or another (local agents or remote cloud agents), with 68% using them daily.” |
The AI Engineering Report 2026 - The AI Acceleration Whiplash
Per developer: task throughput +33.7%; epics +66%; PR merge rate +16.2%; AI acceptance 20%->60%. Per organisation: code churn +861%; bugs/dev +54% (9% in the 2025 cut); incidents +57.9%; incidents-to-PR ratio +242.7%. Median time in review +441.5%; time to first review +156.6%; average time in code review +199.6%; PRs merged without any review, human or agentic, +31.3%.
| Population | 22,000 developers across 4,000+ teams, two years of platform telemetry |
|---|---|
| Method | Observed telemetry with ML/causal analysis comparing each organization lowest- vs highest-AI-adoption periods. Not self-report. |
| Key limitation | MANDATORY CAVEAT: Faros AI sells the engineering-analytics platform this data is drawn from and has a direct commercial interest in the finding that verification is broken. Must never be presented as independent. Customer base is self-selected. Causal identification strategy is not independently auditable. IMPORTANT: the headline throughput figures are normalised PER DEVELOPER and are not organisational delivery measures. The flat-organisational-DORA finding belongs to the 2025 edition of this report, a different dataset (10,000 developers, 1,255 teams), and must not be attributed to this one, which states the opposite: 'these numbers represent real delivery acceleration'. |
| Commercial interest | Faros AI sells the engineering-analytics platform this telemetry comes from. |
| Licence | Proprietary - no licence declared |
| Locator | https://www.faros.ai/blog/ai-acceleration-whiplash-takeaways · archived |
| Supporting sentence | “Epics completed per developer are up 66%. Task throughput per developer is up 33.7%. ... Pull requests merged without any review, human or agentic, are up 31.3%. ... Median time in review is up 441.5%.” |
State of Software Delivery 2026
Average daily workflow runs +59% YoY (largest jump in seven years), which the authors warn is misleading alone: median team +4%, bottom quartile no measurable increase. Median team feature-branch throughput +15%, main-branch throughput -7%. Main-branch success rate 70.8%, lowest in over five years, against the authors' own recommended 90%.
| Population | ~28 million CI workflow runs, 22,000+ organizations, 149 countries |
|---|---|
| Method | Observed CI telemetry, seven years of comparable history. Not survey. |
| Key limitation | CircleCI sells CI tooling, though the finding does not obviously favour them. Population is CircleCI customers, not all software organizations. Workflow runs are a proxy for delivery activity, not for value. |
| Commercial interest | CircleCI sells continuous integration and pipeline validation; the report is sponsored by Thoughtworks. |
| Licence | Proprietary - no licence declared |
| Locator | https://circleci.com/blog/five-takeaways-2026-software-delivery-report · archived |
| Supporting sentence | “the average number of daily workflow runs increased 59% year over year ... But that number is misleading on its own. The median team? Up just 4%, while the bottom quartile saw no measurable increase at all. ... For the median team, feature branch throughput increased 15%, but on the main branch, throughput fell by 7%. ... Main branch success rates dropped to 70.8%, the lowest in over five years and well below CircleCI's recommended benchmark of 90%.” |
State of Code 2026 developer survey
96% do not fully trust AI-generated code correctness; only 48% always verify before committing. AI = 42% of committed code, expected 65% by 2027. 64% use autonomous agents; 72% of AI users use it daily. 38% find reviewing AI code more effortful than human-written code; 53% report code that appears correct but is unreliable; 35% access AI tools via personal accounts; 88% report some negative impact.
| Population | 1,100+ developers globally |
|---|---|
| Method | Survey. Fielding dates, sampling method and margin of error NOT disclosed. |
| Key limitation | Two independent problems: (1) Sonar sells code-review and code-quality tooling and benefits commercially from this finding; (2) methodology disclosure is thin - no fielding window, no sampling method, no margin of error, no regional breakdown. Do not headline without independent corroboration. |
| Commercial interest | Sonar sells code-review and code-quality tooling. |
| Licence | All rights reserved |
| Locator | https://www.sonarsource.com/company/press-releases/sonar-data-reveals-critical-verification-gap-in-ai-coding/ · archived |
| Supporting sentence | “"96% Don't Fully Trust Output, Yet Only 48% Verify It"; "AI accounts for 42% of all committed code - a volume that developers expect will rise to 65% by 2027"; "64% of developers have started to use autonomous AI agents"; "38% of developers noting that reviewing AI-generated code" is more effortful; "35% report accessing AI coding tools via personal accounts".” |
Adoption and Impact of Command-Line AI Coding Agents: A Study of Microsoft Early 2026 Rollout of Claude Code and GitHub Copilot CLI
Adopters merged ~24% more pull requests than they otherwise would have; lift sustained across the entire four-month observation window. Adoption spread through peer networks; retention correlated with coding activity rather than demographics.
| Population | Tens of thousands of engineers at Microsoft |
|---|---|
| Method | Quasi-experimental observational study of an organization-wide rollout, four-month window, early 2026 |
| Key limitation | Preprint, not peer reviewed. Single company, and that company co-owns GitHub Copilot. Merged pull requests are a throughput proxy, not validated value - the same metric Faros shows rising while delivery stays flat. |
| Commercial interest | Authors are Microsoft employees; Microsoft sells the tools under study. |
| Licence | CC BY 4.0 |
| Locator | https://doi.org/10.48550/arXiv.2607.01418 · archived |
| Supporting sentence | “adopters merged roughly 24% more pull requests than they would have otherwise. We use merged pull requests as our proxy for output -- acknowledging that a merged PR is not the same as the value it delivers -- and the lift persists across our four-month window.” |
When Code Becomes Abundant: Redefining Software Engineering Around Orchestration and Verification
No quantitative results; position paper.
| Population | n/a |
|---|---|
| Method | Position paper |
| Key limitation | Conceptual, not empirical. MUST-CITE: this is peer-reviewed prior art for this report thesis. Accepted at ICSE 2026, Future of Software Engineering track. Failing to cite it is the fastest way for an academic reviewer to dismiss the report. |
| Licence | CC BY 4.0 |
| Locator | https://doi.org/10.48550/arXiv.2602.04830 · archived |
| Supporting sentence | “SE must redefine itself around human discernment-intent articulation, architectural control, and verification-rather than code construction.” |
Rethinking Software Engineering for Agentic AI Systems
No quantitative results; structured literature synthesis.
| Population | Relevant literature and emerging industry perspectives |
|---|---|
| Method | Structured literature synthesis |
| Key limitation | Conceptual synthesis, not empirical; no stated inclusion protocol. MUST-CITE as prior art for the thesis. |
| Licence | CC BY 4.0 |
| Locator | https://doi.org/10.48550/arXiv.2604.10599 · archived |
| Supporting sentence | “Code is transitioning from a scarce, carefully crafted artifact to an abundant and increasingly disposable commodity.” |
Cheap Code, Costly Judgment: A Case Study on Governable Agentic Software Engineering
12-week first-person case study; 88 field notes; 420 KLOC production code; 1.16 MLOC of tests, lints, documentation and tooling.
| Population | One expert engineer, one system (document accessibility remediation) |
|---|---|
| Method | First-person longitudinal case study |
| Key limitation | n=1 engineer, single system, self-observed. Cannot generalize. MUST-CITE: closest published work to this report proposed operating model, and shares the two-clause title construction that ruled out one of our candidate titles. |
| Licence | arXiv.org perpetual non-exclusive license 1.0 |
| Locator | https://doi.org/10.48550/arXiv.2607.01087 · archived |
| Supporting sentence | “In contrast to existing governance models that derive controls from known obligations, governance conversion explains how controls are discovered from failures that become visible only during agentic work.” |
Prior use of terms this report uses: context debt, intent debt, cognitive debt, comprehension debt, intent engineering, verification debt
| Method | Terminology prior-art search |
|---|---|
| Key limitation | PRIOR-ART RECORD, not evidence for any claim. Context Debt, Agent Legibility and Intent Engineering are already in use by others; Evidence Engineering collides with Evidence-Based Software Engineering (Kitchenham et al., c.2004); Cost per Validated Outcome near-collides with Cost Per Outcome. AWS CTO Werner Vogels popularised 'verification debt'. All are credited as prior art, not coined here. |
| Licence | CC BY 4.0 |
| Locator | https://doi.org/10.48550/arXiv.2603.22106 · archived |
| Supporting sentence | “Cognitive debt is a team-level, project-level property reflecting the erosion of shared understanding across a software system over time ... intent debt, the absence of externalized rationale that developers and AI agents need to work safely with code.” |
Uplift update: re-running the developer productivity study
57 developers, 143 repositories, 800+ tasks. Returning participants: speedup of -18% (CI -38% to +9%). Newly recruited developers: speedup of -4% (CI -15% to +9%). Original study ran Feb-Jun 2025; the new experiment began August 2025.
| Population | Experienced open-source developers; returning participants plus a newly recruited cohort |
|---|---|
| Method | Randomized controlled trial, follow-up to S03 |
| Key limitation | METR states that significant selection bias undermines the conclusion: developers increasingly declined to participate rather than work without AI assistance. Both confidence intervals cross zero, so neither result excludes no effect. |
| Licence | All rights reserved |
| Locator | https://metr.org/blog/2026-02-24-uplift-update/ |
| Supporting sentence | “Based on conversations with study participants, we believe it is likely that developers are more sped up from AI tools now — in early 2026 — compared to our estimates from early 2025.” |
If you quote a statistic from the report, cite the underlying source, not the report. The locator above is the origin, not a restatement of it. If a figure here does not match what you find at the locator, that is an error worth reporting on the errata page.