Research
Original data on AI-native software delivery
Surveys, censuses of public repositories, telemetry studies, and benchmarks from Stride and the broader software-delivery community. Methodology is always published; sample sizes are disclosed; raw datasets are linked when public. Cite freely.
- N=7,370
What 7,370 popular GitHub repositories put in AGENTS.md and CLAUDE.md
33.4% of the 7,370 active GitHub repositories with 5,000 or more stars carry an instruction file for coding agents, and AGENTS.md is the most common (25.3%), ahead of CLAUDE.md (19.2%). The median AGENTS.md runs 819 words, more than double the 200 to 400 that Codex's /init command asks for, and 10.6% state a rule against leaking secrets. Of the 1,013 repositories with both files, 25.9% leave AGENTS.md out of what Claude Code loads by default.Read Engineering Burnout & Process Debt 2026
Engineering burnout has risen across the 2019–2024 window (39%→62%→53% in the Yerbo cycle; 41% Stack Overflow 2024). Almost no engineering survey uses the validated MBI. The literature points strongly at autonomy and process fit; process debt is rarely measured. Volume 0 reads the clinical literature; Volume 1 (Q3 2026) lands the 600-engineer study.ReadDORA Metrics in Practice 2026
Seven years of DORA Accelerate data converge on four metrics that predict org performance, and one repeated misinterpretation of what they prove. DORA explicitly does NOT claim causation; the popular reading routinely ignores that. Volume 0 reads the literature; Volume 1 (Q4 2026) lands the AI-era segmented refresh.ReadSprint Estimation Reality 2026
Software estimates have been ~30% optimistic on average for 43 years (Halkjelsvik & Jørgensen 2012). The 2026 question isn't whether AI fixes that. It's whether AI hides it. Volume 0 synthesises the literature and pre-registers Stride's 500-person calibration study; Volume 1 lands at this URL in Q4 2026.ReadState of AI Software Delivery 2026
AI adoption in software delivery is close to universal: 90% of DORA's 2025 respondents use AI at work, and 84% of Stack Overflow's 2025 respondents use or plan to use AI tools. Productivity findings still disagree: a 2023 Copilot experiment measured 55.8% faster completion of a standardised task, METR's 2025 trial measured 19% slower on real open-source work, and METR now calls that result out of date, with newer data pointing to small gains. Volume 1, the Stride primary study, has not been published yet.Read
Methodology & press materials
- What 7,370 popular GitHub repositories put in AGENTS.md and CLAUDE.md:Methodology
- Engineering Burnout & Process Debt 2026:MethodologyPress kit
- DORA Metrics in Practice 2026:MethodologyPress kit
- Sprint Estimation Reality 2026:MethodologyPress kit
- State of AI Software Delivery 2026:MethodologyPress kit
About Stride Research
- What is Stride Research?
- A program of original surveys, telemetry studies, and landscape syntheses on AI-native software delivery. Reports are published with full methodology disclosure (sample sizes, recruitment, statistical methods) so journalists, peer reviewers, and engineering leaders can assess the rigor before citing.
- Can I cite Stride Research in my publication?
- Yes. Reports are licensed CC BY 4.0. Attribute with the report title and a link to the source URL. Where datasets are released (look for the "Download raw data" link), reproducibility notebooks are linked from each report's methodology page.
- How does Stride collect research data?
- Four sources: pre-registered surveys (recruited via Prolific Academic, weighted to engineering-leader population), telemetry from Stride beta-customer workspaces (anonymised, opt-in), synthesis of existing public studies (DORA, METR, Octoverse, Maslach Burnout Inventory), and censuses of public data, such as the scan of public GitHub repositories behind the AGENTS.md and CLAUDE.md study. Every numeric claim in our reports is attributable to one of these.
- Are the surveys peer-reviewed?
- Surveys are pre-registered via OSF before fielding (the methodology page links the registration). They are not formally peer-reviewed in the academic sense, but the pre-registration plus full instrument disclosure plus dataset release makes them auditable in a way most industry research is not.
- When will the next report ship?
- Volume 0 (landscape synthesis) of each report is live now. Volume 1 (primary findings from the pre-registered surveys) is planned for the DORA refresh and Sprint Estimation in Q4 2026; the Engineering Burnout and State of AI Software Delivery volumes, planned for Q3 and July 2026, have not been published yet. Each Volume 1 replaces the synthesis sections of Volume 0 at the same URL.