Linear published its first data report this month, and it is the most useful thing anyone has put out on AI and software teams in a while — because it does not measure tokens.
Model companies publish token counts. Coding-tool companies publish lines generated. Linear sits at the other end of the pipe: tens of thousands of teams run the whole workflow through it, from the first issue to the pull request that closes it. So the report, written by head of data Tim Qi and covering paid workspaces from June 2024 to August 2026, can ask a question the token charts cannot. Teams are producing more. Are they finishing sooner?
They are not.
The output numbers are real
Pull requests opened per paid workspace are up 111% against a June 2024 baseline, across 47,900 workspaces. The line was flat for the first year and bent sharply upward from February 2026 onward.
Almost all of that comes from one cohort. In a fixed group of 6,887 paid teams — 4,280 with a coding agent connected, 2,607 without — the agent teams went from 21 pull requests a week to 65. The teams without an agent went from 8 to 10. Agent teams were already higher-output before agents existed, so the levels are not directly comparable, but each cohort against its own baseline is clean: roughly a tripling on one side, noise on the other.
Issue creation moved further. Two years ago fewer than one issue in a thousand inside Linear was written by AI. In the week of 3 August 2026, agents and MCP clients created 2.435 million issues against 2.481 million from people and integrations. On the current slope, machines will be writing more of the backlog than humans within months.
And the people writing code changed. The share of product managers attaching a pull request went from 3% to 10% over two years; designers from 1% to 8%; founders from 11% to 23%. Linear only counts pull requests in repositories connected to it, so those are floors.
The time didn't come back
Here is the part worth printing out.
Comparing June 2025 with June 2026, time spent creating and triaging issues rose for nearly every function — engineering up about 17% on create and triage alone, with commenting up five minutes a user per month. Planning time — customer requests, docs, projects — barely moved at all. And two categories of work that did not exist a year ago, chatting with AI and delegating issues to agents, now show up in every function's month: product added five minutes on AI chat, engineering two, founders four.
Nothing shrank to make room. AI landed on top of the existing job.
Qi's own summary: "teams are working more, not less, suggesting AI has a Jevons paradox quality beyond token consumption."
Our take: Every AI business case written in the last two years assumed the same shape — same output, fewer hours, cost comes out. Linear's data says the opposite shape is what actually happens: more output, same hours, cost stays. That is not a failure, it is a different trade, and it is a much harder one to sell to a CFO who was promised headcount savings. If your 2026 plan has a line item for time saved, this report says find a different justification before someone audits it.
What the report can't tell you
Linear is careful about this and so should you be. It counts pull requests opened, not merged — an opened PR says nothing about whether the change was any good. It sees only paid Linear workspaces, which skews toward fast-moving software companies. Any AI usage outside Linear is invisible. And as Qi concedes, "looking at pull requests indicates motion rather than value."
That caveat cuts both ways. It also means the 111% could be measuring a wave of small mechanical changes as easily as real shipped product.
What to watch
- Whether merged rate holds. Linear says future editions will trace the full lifecycle now that code review runs through the product. Opened-to-merged is the number that turns motion into value.
- The review queue. If PRs triple and reviewers don't, the bottleneck simply relocates from writing code to approving it — and review is where the time went in the create-and-comment data.
- Executive adoption sticking. CEOs at companies of 201-plus went from 9% to 36% AI-active in six months, the biggest jump in the report. Curiosity spikes look identical to durable change for about two quarters.
- Non-engineer PRs. Product managers tripled their share in two years. Watch whether that becomes normal or gets walled off after the first bad incident.