AI Coding Agents Generate More Code, Not More Software

What happened

A new study examining the real-world impact of AI coding agents has found that while these tools generate significantly more code, that increase is not translating into a proportional rise in shipped software. According to reporting from Ars Technica, researchers found that the efficiency gains from AI-assisted coding are effectively "absorbed" elsewhere in the development pipeline — specifically at the human code review stage, which remains a persistent bottleneck.

In other words, developers using AI agents can produce code faster than before, but the downstream process of reviewing, validating and approving that code has not sped up at the same rate. The result is a mismatch: more raw output, but not more finished, deployable software.

Why it matters

The finding complicates a widely held assumption in software and digital-transformation circles — that AI coding assistants directly translate into faster product delivery. It suggests that throughput in complex, multi-stage workflows is governed by whichever stage is slowest, not by whichever stage has been automated most aggressively. Speeding up code generation without addressing review capacity simply shifts the constraint rather than removing it.

For organisations investing in AI-assisted development, this points to a need to look at the entire delivery pipeline — planning, coding, review, testing, deployment — rather than optimising a single stage in isolation. It also raises questions about how review processes themselves might need to evolve, whether through AI-assisted review, revised quality gates, or changes to how teams are structured and incentivised.

The Renascence take

This is a classic case of automating the visible task while ignoring the invisible constraint — and it is not unique to software engineering.

Organisations love to measure the part of a process that is easiest to speed up, not the part that actually determines overall throughput. Code review here plays the same role that a slow contact-centre approval step, a compliance sign-off, or a manual QA check plays in a customer journey: it is the quiet bottleneck nobody budgets for until automation elsewhere floods it with extra volume. A customer-obsessed operator should treat AI adoption as a systems redesign exercise, not a point upgrade — map the whole value chain first, identify the true constraint, and only then decide where automation actually pays off. Otherwise, faster input simply becomes a bigger backlog.