AI coding agents make engineers write more code, but not ship more software. A Harvard study of 718 firms found that after companies adopted agents, lines of code per engineer rose 30% and pull requests 23%, yet resolved Jira issues showed no statistically significant increase. The reason: pull-request review took 49% longer.
The paper, "Artificial Intelligence in the Firm: Bottlenecks in Software Production" by Fiona Chen and James Stratton (current version 4 August 2026), drew wide coverage in October 2026, including Ars Technica. It is a job-market paper, not yet peer-reviewed, so treat it as strong early evidence rather than settled fact. Below is what it measured, what it means if you are budgeting an MVP or a product roadmap, and why regulated teams feel the review bottleneck first.
Agents raised coding activity sharply, but the extra code did not turn into proportionally more finished work. The authors used a staggered difference-in-differences design on data from Jellyfish, an engineering analytics platform: about 300 million work events from roughly 700,000 workers at 718 firms, January 2021 to March 2026, covering GitHub Copilot, Cursor and Claude Code.
| Measure (per worker-month) | AI coding assistants | AI coding agents |
|---|---|---|
| Lines of code added | +12% (not significant) | +30% (baseline 5,017) |
| Commits | +9% | +20% |
| Pull requests | +5% (not significant) | +23% |
| Resolved Jira issues and epics | Small, not significant | Small, not significant; gains above 12% ruled out |
| Time from PR submitted to merged | No significant effect | +49% (baseline 7.03 days) |
| Comments per pull request | — | +35% |
Source: Chen & Stratton (2026), figures as stated in the paper. Adoption was fast: the share of firms using any AI coding agent went from near zero in October 2024 to over 95% by January 2026.
Because writing code is a minority of an engineer's week, and the step after it got slower. The paper estimates engineers spend only 30–42% of their time writing code (30% in a survey of 100 engineers, 42% imputed from the platform data). The rest goes to planning, review and coordination, and agents mostly speed up the typing part.
A quick illustration shows why the ceiling is low. If coding is 30% of the work and an agent makes coding 30% faster, the whole cycle gets about 7% shorter, and that is before review slows down. Our arithmetic, not the paper's, but it matches the paper's result: a large gain at the keyboard, a small or zero gain at the release.
The study found the review step absorbing the difference. After agent adoption the share of pull requests with changes requested nearly doubled, comments per PR rose 35%, and the share of workers doing code review rose 14%. Firms moved people into review rather than shipping more.
Not yet. By March 2026 nearly 80% of firms in the data had used AI tools in code review, but only 23.3% of review comments were AI-generated and only 10.8% of pull requests received at least one AI comment. The bottleneck persisted after firms adopted AI review tools, because humans still carry the decision to merge.
That is the right call in most codebases, and a requirement in some. PCI DSS 4.0.1, for example, expects bespoke software to be reviewed before release, with manual reviews done by someone other than the author (our plain-English guide to PCI DSS 4.0). In HIPAA and SOC 2 environments, change management evidence is what auditors ask for. An agent can draft the change; a person still has to own it.
Not in this data. The authors estimate a precise zero effect on overall employment, ruling out a decline larger than 2.9%, and a noisier zero for engineering employment, ruling out a decline larger than 13.8%. The productivity gain did not pass through to output, and it did not pass through to layoffs either.
Be sceptical of any quote that discounts a build because "AI writes the code". The study says the expensive part of software is not typing; it is deciding what to build, reviewing it and getting it safely into production. Five practical consequences:
Treat the agent as a fast, untrusted contributor and design the pipeline around review. That is the same rule we apply to agents inside products. In practice:
Choosing a tool is the smaller decision. Our comparison of AI dev agent platforms covers where each one runs code and how to cap its access; our AI agent security checklist covers the controls.
No. It is a job-market working paper (first version 7 January 2026, current version 4 August 2026). The data partner reviewed it only to check that no proprietary information was disclosed, not for scientific content.
GitHub Copilot and Cursor as assistants, and agent features including Cursor Agent and Claude Code, observed through the Jellyfish platform at 718 firms.
At the individual level, yes: 30% more lines of code, 20% more commits and 23% more pull requests per worker-month. At the firm level, the gain in finished work was not statistically significant, because review became the bottleneck.
No. They are useful for scaffolding, tests and migrations. The lesson is to pair them with strong review, small changes and good tests, and not to expect a proportional cut in cost or time.
BeevR is a senior, founder-led studio in Hanoi, Vietnam, building software and AI agents for regulated industries at a fixed price per phase, and you own the code from day one. If you are planning a build and want a timeline based on decisions rather than typing speed, see our fixed-price MVP packages or tell us what you are building.