qvib.pro
RU

The productivity reality check (METR)

What for: a sober counter-trend - an RCT showed experienced devs were 19% slower with AI on familiar code, while believing the opposite.

любой

METR RCT (10.07.2025): опытные OSS-разработчики на 19% медленнее с ИИ, при этом верили в +20% ускорение checked 2026-06-01

Updated: 02.07.2026

The productivity reality check (METR)

What it is

Against the backdrop of the agent hype, an important counter-trend emerged: a sober check on productivity. A randomized controlled trial (RCT) by METR showed that experienced open-source developers working on mature projects they knew well were 19% slower with AI tools than without them - even though they themselves were convinced AI had sped them up.

Where it came from

  • 10 July 2025 - METR publishes "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity" (an RCT - the gold standard of evidence).
  • 16 experienced developers, 246 tasks from their own repositories, with AI randomly allowed or forbidden (Cursor Pro + Claude 3.5/3.7 Sonnet).
  • The result clashed with intuition and set off a wave of debate about whether AI actually speeds you up.

Why it matters

  • The perception gap: developers expected a 24% speedup and estimated 20% afterwards - while the real number was -19%. Subjectively "faster" is not objectively faster.
  • Context matters: the effect was measured on mature projects with high standards, where the author is already deep in the material - which is exactly where AI helps least.
  • It's a healthy counterweight to marketing: "a 10x engineer thanks to AI" isn't a universal truth.

How to apply it now

  1. Measure your own productivity with facts (time, throughput), not with the feeling that it "seems faster".
  2. Use AI where the win is real: unfamiliar code or stack, boilerplate, prototypes - not on code you know by heart.
  3. Learn to use the tool: METR named weak prompts and unfamiliarity with the interface among the causes of the slowdown.

What to watch out for

  • This is a snapshot of early 2025 on a narrow group (experienced OSS maintainers) - it doesn't cancel out the gains for beginners and it isn't a verdict on agents; don't over-generalize.
  • Tools have improved a lot since then - rerun the measurement on current models.
  • The takeaway isn't "AI is useless" but "the effect is context-dependent and easily overestimated" - keep that in mind when you evaluate ROI.

Читать по-русски →