AI writes half your issues. Code review is the bottleneck.

Two datasets landed this year that say the same uncomfortable thing from opposite directions. Coding agents made writing code close to free. They did not make shipping software faster, because code review is the bottleneck now, and nobody budgeted for it. If you're paying for agent seats and wondering why the calendar hasn't moved, this is the arithmetic.

What actually happened

Linear published its first data report on AI usage inside its own product, covering January through June 2026. The headline: AI now authors nearly half of all issues created in Linear. Agent-generated issues went from roughly zero in mid-2024 to about 2,435 per week by August 2026 — against roughly 2,500 per week authored by humans, a number that stayed flat the whole time. Agents didn't take work off the queue. They added a second queue next to it.

Output followed. Pull requests are up 111% over two years. Teams using coding agents went from 21 PRs a week to 65. Teams without agents stayed flat at 8 to 10. And the part that should stop you: Linear found no reduction in time spent. Engineering creation and triage time went up about five minutes a month, and chatting with AI added another two to five. Linear's framing is that AI showed up as a new layer of work rather than a replacement for an existing one.

Now the other end. LinearB's 2026 Software Engineering Benchmarks, drawn from more than 8.1 million pull requests across roughly 4,800 teams in 42 countries, measured what happens to those PRs. AI-generated pull requests wait 4.6x longer to get picked up for review than human-written ones. Once someone starts, they review 2x faster. And the acceptance rate for AI-generated PRs is 32.7%, against 84.4% for manual code.

Read those two numbers together. Roughly two out of three AI pull requests don't make it, and they sit in the queue nearly five times as long before anyone finds that out.

Why code review is the bottleneck for your business

You did not buy throughput, you bought inventory. Tripling PR volume with a 33% acceptance rate means you tripled the work-in-progress your reviewers have to wade through to find the third of it that's good. That's a classic queue problem, and adding more agents makes it strictly worse. Before you buy more seats, measure your review capacity in PRs per week and treat it as the hard constraint it is.

Cheap generation raises the value of a good spec. The 4.6x pickup delay isn't laziness — reviewers deprioritize PRs where they have to reverse-engineer the intent. An agent handed a two-line ticket produces a diff nobody wants to open. An agent handed a spec with acceptance criteria produces one someone can approve in ten minutes. The work moved upstream, into defining the thing precisely. That's the same work good engineers were always doing, just now it's the whole job.

Automate the review floor, not the review. Everything a machine can decide should be decided before a human opens the diff: types, lint, test coverage on changed lines, migration checks, dependency diffs, a build that actually runs. Every one of those checks that fires in CI is a rejection that never consumed a reviewer's afternoon. We run agents hard, and the reason it works is that the gate in front of them is expensive to pass.

Watch the metric that actually moved. Not lines shipped, not PRs opened. Track PR pickup time and acceptance rate, split by AI-assisted versus human-authored. If your AI acceptance rate sits near LinearB's 32.7%, your agent configuration is the problem — not your reviewers.

Sources: Linear's 2026 AI usage report, LinearB 2026 Software Engineering Benchmarks.