Typing code was never the bottleneck, and a Harvard study backs that up
AI agents made engineers write about 30% more code, but finished work stayed flat and reviews slowed. Teams that don't fund review are shipping risk they can't see.
AI-generated image
What if the thing AI coding agents speed up is the one thing that never limited your team? A Harvard working paper suggests that may be exactly the situation many engineering organizations are in. They bought a faster typist, and the work piled up at the next desk.
What the paper actually found
According to ComplexDiscovery's summary, Harvard Ph.D. candidates Fiona Chen and James Stratton studied 718 firms. At companies that adopted AI coding agents, engineers added about 30 percent more code. The authors note that this captures the added effect of agents at firms that already had AI coding assistants. Finished work, measured as Jira issues deployed to production, did not measurably rise. Time from pull request submission to merge went up 49 percent, from a 7.03-day baseline.
Two caveats matter. It is a working paper, and the version reviewed shows no sign of completed peer review. The data also comes from Jellyfish, a vendor that sells engineering analytics. The authors say any hidden bias would more likely flatter AI than hurt it, which in our view makes the flat output result harder to dismiss. Still, treat it as strong evidence, not settled science.
Writing code was the cheap part
We think the headline number isn't the 30 percent. It's the zero next to it. Software delivery is a pipeline, and speeding up one stage only moves the queue to the next. Drafting code was probably never what decided how fast a team shipped. Understanding the change, judging whether it's right and taking responsibility for it were. Those jobs sit with humans, and agents haven't removed them. By producing more candidate changes, agents appear to have made them heavier.
The researchers also checked the comforting explanation: that engineers simply moved to bigger tasks. They found no shift in average issue size, and pull requests didn't grow larger either. So the extra volume didn't hide real output. It turned into more things waiting for a person to read.
The risk is the review you quietly skip
Here is where we get opinionated. A 49 percent slowdown creates pressure, and organizations tend to relieve pressure at the cheapest point. Reviewers skim. Approvals get rubber-stamped. Someone proposes using another agent to review the first agent's work. Any of these can make the dashboard look healthy while unreviewed risk goes into production. The speed metric is green and the quality metric is invisible.
That's the real danger of measuring AI by lines written or commits made. Those numbers rose in this study (commits by 20 percent, pull requests by 23 percent) while delivered work didn't. A team that reports on activity will conclude the tools are working. A team that reports on shipped, reviewed work may conclude something less flattering.
What teams should do now
First, change what you measure. Track review time, revision rates and what actually reaches production, not just code produced. ComplexDiscovery makes a similar point for legal buyers: look at approved work product and reviewer time alongside drafting output.
Second, budget for review as a first-class cost. If agents multiply the drafts, reviewers, tooling and protected time have to grow with them. Otherwise you've only moved the queue.
Third, keep a named human accountable for every merge. Whoever approves the change owns it, however it was written.
Our verdict: agents are not a mirage, but the promised productivity gain belongs on the review side of the ledger. Teams that treat reviewing as the scarce skill will turn extra code into extra product. The rest will have a lot more code and the same amount of finished work. They'll also carry hidden risk that surfaces later.
No topics yet: start the first one.
More stories
Claude Code's email-in-User-Agent leak is a default-privacy failure
A coding agent that quietly puts your account email in web requests isn't a quirky bug. It shows why identifiers should be opt-in and why vendors must say what agents transmit.
Postman's silent AI plugins break the one rule dev tools can't
If reports are right, a desktop client that edits your Codex and Claude Code configs unasked has crossed a line. AI features must be opt-in, and API collections need secret-grade care.
Cutting junior developers is a loan against 2036, and the bill is coming
Companies that skip entry-level hires because AI codes faster are saving money this quarter and quietly eating the pool of senior engineers they'll need in ten years.
Linux 7.4's packet pacing offload is boring, and that's the point
Hardware pacing offload won't make headlines, but it shows where kernel performance work is heading: away from flashy features and into the plumbing between CPU and NIC.



