DORA Metrics in the AI Era

Narrative

Independent commentary on the DORA metrics largely treats the annual report itself as a text to be parsed rather than a settled scoreboard. RedMonk analyst Rachel Stephens' close reading of the 2024 Accelerate State of DevOps report is the most cited independent analysis: she flags that DORA reclassified time to restore as a throughput measure and added rework rate to stability, and floats a theory-of-constraints reading that "writing code isn't the bottleneck to deploying reliable applications." Her 2025 follow-up tracks the reversal that made the throughput sign flip positive while noting DORA's own language that AI adoption "not only fails to fix instability, it is currently associated with increasing instability," and records that in 2025 usage rose to 90% of respondents at a median two hours per day.

On the METR randomised controlled trial, independent bloggers treat the 16-developer, 246-task sample with more care than headline aggregators. A LessWrong analysis (cross-posted from the "Don't Worry About the Vase" circle) argues the slowdown is real but may not generalise beyond experienced developers working in codebases they already know deeply, citing Emmett Shear's point that general LLM experience does not transfer cleanly to tool-specific fluency with Cursor. The skeptical outlet Pivot to AI frames the result bluntly, noting developers predicted a 24% speedup, felt 20% faster, and were measured as 19% slower, and highlights that "over half the AI suggestions were not usable." METR's own Substack update in February 2026 is notable for methodological candour: a follow-up study begun in August 2025 was abandoned as an unreliable signal because too many developers refused to be randomised into the no-AI arm, which the team says "likely biases downwards our estimate of AI-assisted speedup."

Practitioner newsletters occupy the space between DORA and vendor dashboards. Abi Noda and Laura Tacho's GetDX newsletter has tracked DORA methodology changes since before the AI era, including a 2022 piece on how organisations misuse DORA as a vanity metric rather than an improvement lever, and more recently argues DORA's four keys alone under-measure AI-assisted teams because deployment frequency inflates with AI-generated scaffolding while review overhead for AI code often exceeds that for human-written code. Gergely Orosz's Pragmatic Engineer newsletter carried the original DevEx framework interview with Nicole Forsgren, Margaret-Anne Storey, Michaela Greiler and Abi Noda, which explicitly frames DevEx as filling the gap "before and after" the commit-to-release window DORA measures. Independent engineering blogs (Jens Oliver Meiert, personal blog) are candid that DevEx and DX Core 4 remain thin on real-world adoption data compared with DORA's decade of longitudinal survey history.

GitClear's code-churn findings are widely relayed by independent technical bloggers (Rob Bowley, jonas.rs) as a corroborating signal for DORA's instability findings, though all treat it as vendor telemetry from a single company's own review-tool customer base rather than an independent academic result. Simon Willison's "vibe engineering" post, arguably the most-cited independent framing of 2025 outside DORA itself, arrives at a similar amplifier conclusion through hands-on practice rather than survey data: AI tools multiply the effect of existing engineering discipline (testing, planning, code review, documentation) rather than substituting for it, a framing several practitioner blogs explicitly map onto DORA's "AI is an amplifier" language.