Six months ago, a complicated feature meant finding the one engineer who knew that part of the system, waiting for them to have capacity, and then watching it slowly work through the pipeline. Domain expertise lived in people’s heads. If that person was out sick or buried in something else, you waited.
That bottleneck is largely gone now.
With agentic coding tools like Claude Code and Codex, more engineers can jump into unfamiliar parts of a codebase and actually be productive. The barrier to entry for modifying or even building new features has dropped dramatically. This is genuinely good. But the bottlenecks didn’t disappear. They moved.
The Review Problem
Here’s the most obvious one: engineers are producing significantly more code. Not vibe coding, not sloppy work. Careful, deliberate, well-reviewed-before-submission code. They’re just doing more of it. An engineer who used to have one PR open now has three. Multiply that across a team.
The pull request model we’ve relied on for years wasn’t designed for this volume. The process is straightforward and familiar: propose a change, get peer review, address feedback, merge. But when a five-person team is churning through work at three to five times the previous rate, that review queue becomes a serious constraint. And the work isn’t simpler. Major refactors, which these tools make much more feasible to take on, are genuinely hard to review effectively.
Yes, the same AI tools can assist with code review. But that only partially addresses the problem, because review isn’t just about catching bugs. It’s how teams build shared understanding of the system.
The Knowledge Gap
Which brings up the second bottleneck: system-level knowledge. Traditionally, a technical lead or senior product owner could hold the full picture of a system in their head. They knew how pieces fit together, where the bodies were buried, what would break if you changed the wrong thing.
Can those people keep up anymore? When the pace of change accelerates this much, maintaining that holistic understanding becomes a fundamentally different challenge. I think the answer is that every engineer is going to need to develop this skill. The days of taking your little slice of the codebase and staying in your lane are numbered. Understanding how your work fits into the broader system isn’t optional anymore. It’s table stakes.
The Idea Problem
There’s an even less obvious bottleneck forming, and it’s upstream of engineering entirely. When the time to build and ship something shrinks this dramatically, the constraint moves to having good ideas and validating them.
I can see how teams end up in a cycle of endless tweaking. Ship a feature, change it, change it again, never pausing long enough to understand whether the version from two iterations ago was actually working. Because it’s so easy to make changes now, a senior leader can say “actually, shape it like this instead” and it’s done by tomorrow. That speed is a double-edged sword. Our processes for validating whether something is working aren’t built for that cadence.
The Messy Middle
We’re in an uncomfortable in-between space right now. We can see that things are different and getting more different. But our tools, processes, and metrics were all designed for the old pace.
Project tracking doesn’t account for a three-person team producing the output of a ten-person team. QA processes weren’t built for this volume of change. And then there’s the human side: what happens when two engineers on a team are fully invested in these tools and operating at a dramatically higher pace, while two others are working the way they did three years ago? The production gap between those groups is substantial, and it creates tensions that aren’t easy to resolve.
I don’t have neat answers to any of this. The dust hasn’t settled. And I’m positive there are other bottlenecks shifting across the development lifecycle that I’m not even thinking of right now. These are just the ones that are top of mind for me today. The old constraints around implementation complexity and siloed expertise are dissolving fast. The new ones are still taking shape, and I suspect we’ll keep discovering them for a while.
We’re all figuring this out in real time.