You have 1 article left to read this month before you need to register a free LeadDev.com account.
Estimated reading time: 3 minutes
Key takeaways:
- AI is making developers busier, not more productive. A 741% increase in code written translates to just a 20% increase in releases.
- The bottleneck is everything after the code.
- The developer’s job has changed: writing code is no longer the primary skill – evaluating it is.
AI-coding tools have become a vital part of any developers’ toolbox in just a few years. The release of Fable, Anthropic’s latest model, propels the power forward even further.
However, adoption of AI-coding tools hasn’t always been smooth, with software engineers struggling to figure out how to make them fit into their workflow, and their managers grappling with how the job is changing.
A new study analyzing more than 100,000 GitHub developers provides a reality check for engineering managers: writing code is not the same as shipping code.
The researchers from the Massachusetts Institute of Technology and the Wharton School of Business found that, while task-level productivity has skyrocketed, human bottlenecks in the software development lifecycle are compressing these gains – making devs busier, but not necessarily more productive.
“We’re definitely seeing more code being generated, but that doesn’t automatically translate into more value being shipped,” says Alfonso Graziano, AI technical lead at Nearform, a software developer. All it means is that AI can shift where work ends up getting stuck, he says.
Your inbox, upgraded.
Receive weekly engineering insights to level up your leadership approach.
Evolution, not revolution
The researchers evaluated three successive generations of AI tools – autocomplete tools like the original GitHub Copilot, synchronous agents like Claude Code, and async agents that operate autonomously – tracking their impact from early autocomplete features to the latest autonomous agents.
They found that the productivity increase at the granular level of writing code was undeniable. Early autocomplete tools boosted coding activity, as measured by commits, by 40%. Interactive sync agents that code alongside developers were even more effective, pushing the cumulative gain to 140%. Autonomous asynchronous agents raised it to 180%. Those agents increased the number of lines of code developers could write by 741%.
Yet task-level gains “attenuate sharply,” the researchers wrote, as the work moves up the organizational hierarchy. That 741% increase in code written only translates to a 65% bump in pull requests (PRs), and just a 20% increase in actual software releases.
The authors suggest this is down to what they call the “weak-link hypothesis.” AI is effective at generating raw code, but everything after that, from reviewing PRs to integrating changes and managing releases, still rely on human effort and judgment. As human capacity for code review remains largely fixed, the sheer volume of AI-generated code simply creates a traffic jam later in the pipeline.
“A system’s throughput is set by its slowest stage, so if you speed up any other process the throughput does not change,” says Graziano. “The only thing that grows is the pile of work-in-progress in front of the real constraint and if you push hard enough, the system gets worse, because work piles up faster than the constraint can drain it.”
More like this
Managing the change
The bottleneck ultimately extends all the way to the end-consumer. While the researchers observed a moderate spike in new application releases across major platforms like the Apple App Store and Google Play Store since mid-2025, total app usage remained entirely flat.
That, they suggest, means a growing share of these new applications is failing to attract any meaningful audience, proving that flooding the market with AI-assisted software does not automatically translate into value or discovery for users.
If things have changed so significantly without seeing a shift in the end product, how can managers make it so that the benefits of AI support don’t get lost by the wayside?
A mentality shift would be a start, reckons Josh King, CTO of &above, an enterprise AI firm. “Developers still think of themselves as writers of code,” he says, but that’s not correct anymore. “Their actual job is now evaluators of code.”
New York • September 15-16, 2026
Full LDX3 lineup is here 🙌
King explains how &above has gone from three or four open PRs at one time on a project to 20-plus. “Review is now about half of a developer’s job, but many teams still treat it as the final step in the process rather than the primary constraint on delivery,” he says. So the firm has overhauled how it thinks about evaluation and testing of code to capitalize on the changes.
“The question isn’t whether AI-coding agents make developers faster,” King says. “They clearly do. The question is whether organizations can adapt their review, testing, and maintenance processes quickly enough to keep up with the volume they’re creating.”
