AI coding agents produce 30% more code, not 30% more software. The technical debt that results, and how to avoid it.
A study reported by Ars Technica in early October 2026 laid it out clearly: an AI coding agent that writes 30% more lines of code doesn't deliver 30% more software. It's often the opposite. Write speed explodes, human review speed stagnates, and something eventually gives.
This text defends a simple thesis: the real bottleneck for coding agents like Claude Code or GitHub Copilot isn't code generation. It's review. And the technical debt that results is invisible in every velocity dashboard.
What the AI coding agent study actually reveals
Ars Technica, analysis of AI coding agents study, October 9, 2026 "The introduction of AI coding agents at a firm leads to a 30 percent increase in total lines of code generated."
Translation: the arrival of a coding agent in a team bumps up the volume of code produced by roughly a third. So far, nothing surprising, that's exactly the promise vendors sell. The problem starts after, according to Ars Technica.
Forkast, coverage of Harvard AI coding agents study, October 11, 2026 "The first large-scale empirical study of AI coding agents at enterprise scale finds that agents boost code volume by 30%, but the gains don't translate to software output."
More code, but not more software shipped. No more features in production. No more tickets resolved. According to Forkast, reporting conclusions from a study across multiple enterprises that adopted these tools, the limiting factor identified is review capacity, not generation capacity.
That matches what we observe with our clients testing Claude Code or Cursor on existing codebases. We already touched on this in How to Build an AI Agent Without Making the Mistake Everyone Makes: giving an agent broad permissions without guardrails produces speed, not reliability.
Why more code doesn't mean more software shipped
Most engineering dashboards measure generation velocity. Never review velocity. We're looking at the wrong end of the pipe.
A human developer writing code structures it, on average, in a way they know they can defend in review. An AI agent doesn't have that brake. It produces what compiles and what passes the tests you gave it, not what will be easy to maintain in six months. Result: longer, denser pull requests and review cognitive load that climbs faster than the number of developers capable of absorbing it.
And a tired reviewer approves faster. It's human.
The hidden cost: technical debt nobody budgets for
Classic technical debt gets budgeted after the fact: you know you cut a corner, you write it down, you refactor it later. Debt generated by a coding agent is different, it's often invisible when it forms, because the code looks clean. It's well indented, well named, sometimes even commented. It doesn't look like debt.
Take a team of 4 or 5 developers letting an agent generate PRs of several hundred lines on an existing module. A review that took 15 minutes on an equivalent human PR can easily take 45, because the reviewer has to mentally reconstruct the intent behind each choice, intent the agent never formulated, just executed.
This dynamic has a honest limit: it mostly hits teams that let PRs grow. A team already practicing short PRs and tight trunk-based development absorbs the shock much better, agent or not.
The strongest counter-argument: what if agents review too?
The most serious objection to this thesis is that review agents already exist, static analysis automation, auto-generated PR comments, risky pattern detection. If an agent generates and another reviews, where's the problem?
The problem is error correlation. A language model that generates code with a business logic flaw has a good chance of missing exactly the same flaw when asked to review it; both tasks lean on the same understanding, often the same model family. A human reviewer brings cognitive divergence: they don't reason the same way as whoever (or whatever) wrote the code. That's precisely what automated review doesn't yet reliably replicate.
That doesn't mean AI-assisted review is useless, it catches real mechanical bugs, forgotten null-checks, type inconsistencies. But it doesn't replace judgment on architecture and long-term maintainability, the subject we dug into in Claude AI Agent: Why Even the Best Model Isn't Enough in Production.
Conclusion
Three things to take away. First, an AI coding agent that produces more lines doesn't mechanically produce more software value, the bottleneck just moved to review. Second, technical debt generated by an agent is harder to spot than classic debt, because the code has the appearance of quality without always having the substance. Finally, the answer isn't to slow adoption of agents, but to tighten review discipline and PR size before giving them more room to run.
At fstck, we work with teams that want to industrialize AI agents in production without sacrificing the quality of the custom web development they ship to their clients. If your review backlog swelled since the coding agent arrived, now's the time to talk about it.
Frequently asked questions
Does an AI coding agent generate more bugs than a human developer?
Not necessarily more bugs, but different ones, often tied to business intent rather than syntax. The study reported by Ars Technica doesn't measure bug rate directly, but shows that code volume increases without a proportional increase in actual problems solved.
How do you reduce technical debt created by an AI coding agent?
Limit the size of generated pull requests, require the agent to justify its architecture choices in the PR description, and keep a senior human reviewer on any change touching business logic or data.
Should you stop using Claude Code or GitHub Copilot after this study?
No. These tools stay useful for speeding up boilerplate code or tests. The signal from the study isn't "stop," it's "scale your review capacity along with your agent adoption," or risk accumulating invisible debt.
Why does code review become the real bottleneck with AI agents?
Because the volume of code produced grows faster than human capacity to understand and validate it. A reviewer has a cognitive limit per day; an AI agent doesn't, creating a structural imbalance between generation and validation.
Can AI agents do code review instead of humans?
They can catch mechanical errors, typos, null-checks, known risky patterns. But they often share the blind spots of the model that generated the code, making them unreliable for judging long-term maintainability or the merit of architecture decisions.
