AI coding agents produce more code, but not more software

Meta description: Jellyfish data across 700,000 workers shows AI coding gains are offset by longer code review, creating a bottleneck for software engineering output.

Tags: AI code review, Jellyfish, engineering analytics, Harvard University, software productivity

Featured image: abstract technology illustration with code nodes, review checkpoints, and time-series metrics; no real people, company logos, or text.

A recent Harvard University study based on Jellyfish data finds that an AI code review bottleneck has become a significant constraint on the overall efficiency of AI coding tools, offsetting productivity gains from coding assistants and agents. Across more than 700,000 employees at over 700 software development firms, the researchers saw little evidence that companies increased software output or reduced employment after using the tools. The finding matters because it shifts the conversation from code generation speed to the full production pipeline: if draft code arrives faster than humans can inspect it, the saving may disappear before release.

Before the Measure: Generative Code Without Full Trust

Modern AI coding assistants and agents can produce large amounts of functional code quickly, yet programmers who use them already know not to treat every output as accurate. Review remains a major cost because generated code can contain logical errors, unsafe assumptions, or mismatches with existing systems. That human check is a normal part of software engineering, but it also becomes the pressure point when AI raises the volume of drafts moving into a team's workflow.

From 2021 to March of 2026: A Large Workflow Dataset

To test how actual coding practices play out in real organizations, Harvard University researchers Fiona Chen and James Stratton used aggregated analytics data from Jellyfish, which measures the granular output of engineering teams. The dataset covered 300 million individual work events, including commits and pull requests, alongside issue management software data. It spanned more than 700,000 employees at over 700 relevant software development firms from 2021 through March of 2026.

That scope is important. Instead of asking only whether developers wrote more code, the study looked at the sequence of events through which code becomes software: request, revision, comment, review, and downstream production constraints. The result is a picture of AI coding adoption measured through operational behavior, not self-reported sentiment.

The Bottleneck Emerges: Review Time Absorbs Coding Gains

The study's central conclusion is that speed saved during the actual coding phase can be absorbed by later constraints in the production process. In the observed data, review stages became markedly longer. Pull requests showed a higher likelihood of needing revisions, and reviewers posted more comments. Those changes do not simply show that reviewers became stricter; they show that more AI-assisted work is arriving in a form that requires human attention before it can merge.

As a result, firms may not see a clean rise in software output, nor a reduction in employment, even when AI tools make individual coding tasks faster. The bottleneck is not a failure of the tools alone, but a mismatch between generation speed and the capacity of the review layer.

Current Phase: The Constraint Moves Downstream

The current stage of AI coding adoption can be described as fast drafting, slow validation. The Jellyfish-based study suggests that organizations have not yet converted the time saved while writing code into proportionally faster delivery. What to watch next is whether engineering metrics begin to treat review cycle time, revision frequency, and reviewer comment load as core indicators of AI productivity, rather than counting only commits and pull requests.