AI coding's three conflicting productivity numbers are all correct

Software teams have spent 2026 throwing three numbers at each other over whether AI coding tools help or hurt. Which number a team trusts shapes whether it expands AI tooling or hires more reviewers.

Field experiments that count tasks completed report 26% more work finished per week with AI assistance. METR's randomized trials, which clock wall-clock time on developers' own repositories, found a 19% slowdown. Team telemetry, tracking where the work moved, shows code review time up 441%.

A paper reconciling the evidence posted August 20 argues all three numbers are correct, because each instrument measures something different: tasks finished, time elapsed, or where the labor shifted.

Its most interesting claim is also the most hedged. Gains look real on new code and shrink or reverse on codebases with years of accumulated decisions. The authors call this a conjecture, not a finding.

That split matches what I see on my own projects. Fresh code with an agent moves fast; a codebase where I keep re-explaining constraints a human teammate would've picked up in a week goes slower.

My sample size is one person's side projects. That resemblance is a reason to test the claim, but it doesn't prove it.

Track two weeks of agent-assisted work split by codebase age, and measure wall-clock time to merged rather than time to first draft.

The review-time number holds regardless of that split. A separate study of 248,641 GitHub pull requests found agent-authored code increasingly reviewed by other agents. Cross-vendor review, one company's agent reviewing another's code, grew more than a hundredfold between the first and third quarters of 2025.