Coding agents write 30% more code. Shipping did not move

Your coding agent is working. Your shipping rate is not.

That is the finding of a Harvard working paper by Fiona Chen and James Stratton, "Artificial Intelligence in the Firm: Bottlenecks in Software Production," current version dated August 4, 2026. The authors analyzed Jellyfish engineering-analytics data covering about 300 million work events from roughly 700,000 workers at 718 software firms, from January 2021 through March 2026. They compared firms before and after adopting AI coding tools, using the staggered timing of adoption across firms to estimate the effect.

The coding gains are real. After firms adopted fully agentic tools, lines of code rose 30%, commits rose 20%, and pull requests rose 23%. All three effects were statistically significant. Simpler autocomplete-style assistants produced smaller gains: lines of code up 12%, commits up 9%, pull requests up 5%, with only the commits effect reaching statistical significance.

Then the paper looks at what actually shipped. It measures firm output two ways: resolution of Jira issues, which happens once coding work reaches production, and resolution of Jira epics, which represent whole features or projects. For both measures, and for both tool classes, the effects were small, positive, and statistically insignificant. For agents, the estimates rule out an output increase above 12% of the baseline. Code output up 30%. Shipped output up by something the data cannot distinguish from zero, and definitely not 30%.

The paper also checked whether AI adoption changed employment. It did not find significant effects. For agents, the estimate for overall employment was a precise zero, ruling out a decline greater than 2.9%. The engineering-employment estimate was noisier but also rules out a decline greater than 13.8%. Whatever coding agents are doing to headcount at these firms, the data does not show it yet.

So where did the extra code go? Into review.

After agent adoption, the average time to review a pull request rose 49%. The share of pull requests sent back for changes nearly doubled. Comments per pull request rose 35%. Firms responded by shifting labor: the share of workers doing code review grew 14%. And the bottleneck persisted even at firms that had already adopted AI code-review tools, because human reviewers kept the final sign-off.

The authors' explanation is structural, not moral. Writing code is one stage of production. Code must be reviewed and tested before it ships. Their imputation from the data, validated against a survey of 100 engineers, puts writing code at 30 to 40% of an engineer's time. Speed up one stage and the constraint moves to the next one. In their model, AI affects code writing through two channels: it raises coder productivity, and it changes the quality of the draft code entering review. More volume plus more review effort per unit equals a jam at the review stage.

This is the number to carry into your next budget conversation. If you justify agent seats on lines of code, you will overstate the payoff. The paper's prescription is blunt: "Realizing the benefits of AI adoption requires complementary investments in downstream review capacity, rather than simply increasing code output." Buy the seats and the review capacity together, or buy neither. A team that adds agent seats without adding reviewers, or without training agents to produce review-ready diffs, should expect the pattern Chen and Stratton measured: more commits and the same shipped features, with reviewers absorbing the slack.

Three caveats before you quote this in a board deck. First, it is a working paper, not yet peer-reviewed. Second, resolved Jira issues are a proxy for shipped software, not a direct measure. The authors tested one obvious objection, whether AI adoption made issues bigger or more complex so that flat counts hid real growth, and found no evidence of a change in average predicted issue length. Still, a proxy is a proxy. Third, the data ends in March 2026, and the sample is firms that bought an engineering-analytics platform. Newer agents may shift the ratio. The direction of the finding is the durable part. The exact percentages will move.

The finding also reframes the coding-agent price war of the last month. Vendors are competing on per-seat price and per-token cost while the binding constraint sits one stage downstream. A cheaper seat that produces more unreviewed code is not a bargain. It is a longer queue. The same logic applies to the subagent economics now shaping agent design: parallel agents multiply output, and every parallel stream still ends at a human reviewer. Until review capacity scales, the ceiling on shipped software stays where it is.

What remains unknown: whether the bottleneck loosens as agents get better at producing review-ready diffs, whether AI-assisted review ever earns final sign-off authority at serious firms, and whether the pattern holds outside the Jellyfish sample. Watch for Jellyfish, LinearB, or Code Climate to publish their own breakdowns on fresher data. If your org tracks PR review time, you do not need to wait. Compare it before and after your agent rollout. That is the number that decides whether the seats paid off.