What happened
A new study highlights a growing bottleneck for engineering teams using AI coding agents. According to a report from Ars Technica, these tools successfully generate a higher volume of code. However, they do not necessarily accelerate the overall delivery of finished software.
The core issue identified is the human review process. The time required for developers to validate, debug, and integrate AI-generated code is offsetting the gains from faster code creation. This creates a new chokepoint in the development lifecycle.
How the room's reading it
The findings are resonating with engineering leads on forums and X. Many are sharing their own experiences of pull request queues ballooning with AI-generated code that requires meticulous, time-consuming review. The consensus among practitioners is that simply adding an AI agent to an existing workflow isn't a silver bullet for productivity.
Some infra teams see this as a tooling problem — a need for better AI-native review and testing platforms to manage the increased volume. Others argue it's a process problem, suggesting teams need to redefine what they ask agents to do, favouring smaller, more verifiable tasks over entire features.
Sailfish's take
We've seen this firsthand. The problem isn't that the agents are bad — it's that engineering teams are measuring the wrong thing. For years, lines of code was a vanity metric. Now, it's an active impediment to shipping good software. Chasing code volume with AI is like trying to win a race by spinning your wheels faster.
The real task is to reorganise the work itself. We'd push teams to use agents for discovery and scaffolding, not for drafting entire pull requests. If your team's PRs are getting bigger because of AI, you're using it wrong. This is the week to stop measuring code output and start measuring shipped outcomes.