Skip to content

Signals · Judgement Layer

AI coding agents lifted code written by 30 per cent. Finished work stayed flat and review became the bottleneck

A Harvard study of 718 firms found AI coding agents lifted code written by 30% while finished work stayed flat and review took 49% longer.

by ·4 min read·

New here? Start with the free AI Survival Kit →

Part of the guide Your employer is grading your AI use. What to do before the next review

AI coding agents lifted code written by 30 per cent. Finished work stayed flat and review became the bottleneck

ComplexDiscovery reported on 13 October on a Harvard working paper that measures what happens after AI coding agents arrive. Fiona Chen and James Stratton are Ph.D. candidates at Harvard. They tracked 725,938 workers at 718 companies between January 2021 and March 2026, using engineering data from Jellyfish joined to LinkedIn records from Revelio Labs. Engineers at firms that adopted agents added about 30 per cent more code. Finished work did not measurably rise. Getting code through review took 49 per cent longer. The paper is Chen's job-market paper, dated 4 August 2026. ComplexDiscovery found no sign that it has completed peer review, and it studies software only.

The draft got cheaper and the sign-off got slower

Lines of code added rose by 1,494 per worker-month against a baseline of 5,017. Commits rose 20 per cent and pull requests 23 per cent. Then the gain stops. The paper counts an issue as finished when its code is deployed to production. Resolved issues moved by 0.12 per worker-month on a baseline of 3.67, a result statistically indistinguishable from zero. The confidence interval rules out a rise above about 12 per cent of baseline.

The authors checked whether engineers had moved on to bigger tasks. They found no shift in average issue size.

The shortfall sits in review. Time from pull request to merge rose 3.45 days on a 7.03-day baseline. The share of pull requests drawing a formal change request rose about 12 percentage points from a 13 per cent base, nearly doubling. Review comments per pull request rose 35 per cent. The share of workers doing any review rose 14 per cent. The share writing code did not measurably change.

AI review tools did not clear the queue. By March 2026 nearly 80 per cent of the firms had used one. Even so, AI wrote only 23.3 per cent of review comments, and only 10.8 per cent of pull requests received a single AI comment. One junior engineer told the researchers: "After some outages that were caused by AI-generated commits, there were also new rules requiring senior engineer approval for AI-generated code, which created even more review bottlenecks."

What the numbers say about who holds value

Read this as a labour-market signal. Agents sped up the drafting. The work that grew is judgement: deciding whether the work is right, sending it back and putting a name on it. The authors' model gives two causes. More code reaches a fixed pool of reviewers, which moving people into review can largely relieve. The rise in change requests may also signal lower-quality drafts, which the authors say would need better technology.

Employment held steady too. The estimates rule out an overall decline larger than 2.9 per cent. The authors caution that these are short-run effects and could hide job losses at competing firms.

AI does the routine work. You do the thinking. This study puts a number on the second half. ComplexDiscovery applies the same checkpoint to legal work. Harvey's motion-to-dismiss workflow and Cooley's redaction agent both keep a lawyer at the checkpoint, and neither announcement reports how long that step takes.

The risk for an individual is the metric. A team that counts lines written will report a 30 per cent gain that never reached production. If your employer scores AI usage, our guide on how to show impact when your employer grades your AI use explains how to put finished outcomes in front of the people reviewing you.

Your Next Move

  1. Track the review stage on your own work this month: time to approval, rounds of revision and changes requested. That stage decides your output, and few people report it.
  2. Move towards the checkpoint. If one senior person approves everything on your team, offer to take a share of review in your area. The study shows the queue growing, and the person who clears it holds the leverage.
  3. Report finished work in your next review: shipped, approved, filed. Put any figure for AI-assisted volume next to the count that actually made it through.

I build systems like the one publishing this site. → Work with me

About the author

Jo

Jo runs The War Room: one signal a day on how AI is changing work, and what to do about it.

Sources

Get the Briefing

Want more intelligence like this?

The weekly briefing, free — and the AI Survival Kit with it.

Takes 30 seconds. No spam.

More in Signals

See all