AI coding agents are creating a review bottleneck

The bottleneck moved

AI coding agents — Cursor, Claude Code, Kilo, Copilot — are genuinely making engineers faster at writing code. A task that used to take a day might take half a day. But the downstream effect is predictable: more code written means more PRs opened means more reviews needed.

Review capacity hasn't scaled to match. The same three senior engineers who were reviewing PRs before are now reviewing 50% more PRs. The result: PRs sit in the review queue longer, feedback cycles slow down, and the time from "code written" to "code shipped" actually increases. The team is producing more code and shipping it slower.

Review is the new bottleneck

This pattern shows up clearly in the metrics. Teams using AI coding agents see cycle time (PR open to merge) go up, not down, in the first few months. The writing phase shrinks, but the review phase expands to fill the gap and then some.

The problem is asymmetric: AI can help you write code, but reviewing code requires understanding intent, context, and risk. A reviewer needs to know what ticket this PR addresses, what the acceptance criteria are, whether anyone else is touching these files, what Sentry errors might be related, and what feedback from the last round still hasn't been addressed. That context-gathering takes time, and AI coding agents don't help with it.

Context is the missing piece

What if the reviewer didn't have to gather context themselves? What if every PR came with a brief that included the linked ticket with acceptance criteria, a risk assessment based on what files are changing, whether anyone else has open PRs touching the same code, relevant Sentry errors that might be related, and any unresolved feedback from previous review rounds?

That's not a hypothetical. That's what Gary does for every PR. The reviewer opens the PR and the context is already there — the ticket says X, the risk is medium because shared infrastructure is touched, nobody else is in these files, and there's one comment from last round that still needs addressing.

The review itself still requires human judgment. But the 10-15 minutes of context-gathering before the review? That's gone. For a team doing 20+ reviews a week, that's hours saved — and more importantly, faster review turnaround because the friction to start a review just dropped.

Coordination, not more code

The lesson from the AI coding agent wave is that writing code was never the real bottleneck for most teams. Coordination was. Knowing what to work on, who should review it, what context they need, and whether your work conflicts with someone else's — that's where time actually goes.

AI coding agents are valuable. They make the writing phase faster and that's real. But without better coordination, they just move the bottleneck downstream. The teams that will benefit most from AI coding agents are the ones that also invest in coordination tooling: better PR context, smarter reviewer assignment, proactive conflict detection, and sprint health monitoring that keeps the whole pipeline flowing, not just the first stage.