LinearB 2026: agent PRs merge less without owners; AI review lifts yield
Jul 28 LinearB 2.7M-PR benchmarks: elite orgs merge agent PRs at 79% vs 92% human-only (37% at top 60%); ownership — not model skill — is the limiter; AI code review lifts PR yield up to +5–7pts while generation moves the bottleneck downstream.
Source: The AI engineering productivity gap: 2026 benchmarks
LinearB’s 2026 AI benchmarks (published Jul 28; 2.7M PRs, 83k developers, 253 orgs, Feb–May 2026) productively complicate the always-on factory story with merge-rate receipts. Very-high AI users hit 2.3× their June 2025 merge rate by May 2026 while no-AI developers were roughly flat — but autonomous agents are still a thin slice of delivery. At elite orgs, 54% of PRs show AI coding assistance and 45% of merged lines are AI-written, yet under 5% of PRs come from autonomous agents.
The agentic yield gap is the hard number: at the top 10% of orgs, agent-opened PRs merge within 30 days at 79% versus 92% for human-only PRs; at the top 60%, agentic yield falls to 37%. LinearB’s diagnosis is ownership, not capability — an agent PR with no engineer attached tends to sit unmerged (experimentation at the edges, not agents inside delivery). AI code review runs the other way: PRs with an AI review attached merge at the highest rate of any category, lifting yield up to five points over the all-PR baseline and up to seven over human-only at the top 60%. Among very-high AI users, PR yield slipped 4.4% YoY even as merge rates climbed — faster generation moved the bottleneck into review and testing. Token spend at the 90th percentile hits ~$481/dev/month (still under ~4% of fully loaded cost); the useful metric is cost per delivered PR, not raw tokens.
Why it matters for code review
Overnight harnessed momentum only compounds if morning inventory actually merges. LinearB’s receipt is that always-on generation without named owners and review gates produces abandoned agent PRs — verification tax under volume shows up as yield, not as a weaker model. Continuous factories that attach ownership, AI review, and merge policy to every agent PR protect the throughput story; bursts of unowned agent output still lose to a quieter shop that converts opened PRs. FOMO when agents are not working on your behalf is real — so is FOMO when they worked all night and nothing landed.