Most engineering teams adopted AI coding assistants expecting overall delivery velocity to skyrocket. Instead, developers are writing code faster while team-level delivery stays completely flat.
The real grind is that writing code was historically the primary bottleneck in the software development lifecycle. Because of that constraint, engineering teams designed review and approval processes around human-paced authorship. When you introduce AI coding tools, code gets generated rapidly and becomes cheap to change, but the downstream review pipeline remains unchanged. The constraint did not vanish. It just shifted directly into the pull request queue.
Senior engineers who once managed a predictable flow of pull requests are now buried under massive backlogs of machine-generated code. Because adding more human reviewers only treats the symptom, teams need to address the friction upstream before code ever gets generated:
Make specifications explicit from the start: Human engineers historically filled in vague requirements using informal judgment and chat threads. AI agents build exactly what is written, meaning ambiguous specs lead to wasted iterations and faulty code.
Involve non-engineers before implementation begins: Pulling product, design, and security into the review stage made sense when writing code was slow. When code is generated quickly, non-technical stakeholders must help shape definitions before generation starts.
Optimize for the cost of late changes: Stop measuring raw pull request volume or review throughput. The metric that truly matters is how expensive it is to redirect work that started from an unclear or incomplete spec.
When you solve requirement ambiguity upstream, you protect your senior reviewers and ensure that fast code generation turns into shipped software rather than review gridlock.
Read the full article:
Top comments (0)