Linear Watched 6,887 Teams for Two Years. Coding Agents Tripled Their Pull Requests and Gave Back Zero Hours.
Linear published its first data report this week, drawn from its own product telemetry across tens of thousands of paid workspaces. The number everyone will quote: teams with a coding agent connected went from 21 pull requests per week in June 2024 to 65 by June 2026. Teams without one went from 8 to 10.
The number nobody will quote is the one that should change how you plan your week. Across every function Linear measured, time spent on existing work did not go down. Not for engineers, not for founders, not for anyone. The AI work landed on top.
The output numbers, and what they are measuring
The cohort is 6,887 paid teams: 4,280 with a coding agent connected, 2,607 without. Fixed cohort, so this is the same teams tracked over time rather than a shifting sample. Across all paid workspaces, pull requests per team per week are up 111% against a June 2024 baseline, with the curve roughly flat for the first year and bending upward through 2026.
Linear is unusually careful about the caveats, and they matter. They count pull requests opened, not merged. Agent teams were already higher-output before coding agents existed, so 21 versus 8 at baseline means the two cohorts were never comparable in level, only in their own trajectories. And they say plainly that an opened PR says nothing about the value of the change.
That last one deserves more weight than it will get. A tripling of PR volume is compatible with three very different worlds: a team shipping three times the value, a team splitting the same work into three times as many pieces, or a team generating a lot of changes that need review. The data cannot distinguish them. What it does establish is that the motion is real and it is concentrated almost entirely on the agent side.
The part that should worry you
Between June 2025 and June 2026, minutes per user per month on creating, triaging, and commenting went up in nearly every function. Engineering rose from 24 to 28 minutes on create and triage, and 35 to 40 on commenting. Founders rose from 40 to 57 minutes on creation and 39 to 64 on commenting.
Meanwhile, chatting with AI and delegating issues to agents are categories that did not exist a year ago and now appear in every function's week. Product leads at 5 minutes on AI chat, founders at 4 plus another 2 on agent issues.
Nothing shrank to make room. Linear's closing note says it directly: the time spent on product development is going up rather than down, and they reach for Jevons paradox to explain it. Cheaper code generation did not reduce the amount of time spent on code. It increased the amount of code, and the coordination around code scaled with it.
I have felt this exact shape running content and code pipelines solo, and I did not have a name for it until I read this. The generation step got dramatically cheaper. The reviewing, the deciding what to generate, and the cleaning up after did not, and those grew in proportion to output. The bottleneck moved. It did not disappear.
Why this hits a solo operator differently
On a team, the extra coordination is distributed. Someone triages, someone reviews, someone writes the issue. The founder's 26 extra minutes per month on commenting is annoying but absorbable because there are other people absorbing the rest.
Solo, every one of those minutes is yours. You are the agent's product manager, its reviewer, and the person who has to understand the code at 2am when it breaks. Tripling PR volume means tripling the review surface for one person, and review does not parallelize the way generation does.
So the finding I would actually act on is not "connect a coding agent, get 3x output." It is: whatever throughput gain you get, budget for the coordination it creates, because the data says that cost is real and nobody is reporting it as time saved.
The role-blurring finding is the sleeper
The share of product managers attaching a pull request went from 3% in June 2024 to 10% in June 2026. Designers went from 1% to 8%. Founders from 11% to 23%. Engineering itself went from 20% to 34%.
And AI now authors just under half of all issues created in Linear. Two years ago it was fewer than one in a thousand. In the week of August 3, 2026, agents and MCP clients created 2,435 thousand issues against 2,481 thousand from people and integrations. That crossover is weeks away, if it has not happened already.
For anyone running a one-person company, this is the finding with the longest tail. The reason you hire a specialist is that some work requires a skill you do not have. If designers are shipping code and PMs are opening PRs at three times the prior rate, the threshold for "I need to hire someone for this" moved. Not to zero, but it moved, and it moved in the direction that favors staying small.
Where I would push back on the data
Two limitations Linear names, and one they do not.
They name that this only covers AI usage inside Linear, so it is their customer base rather than the market, and that company-size data comes from third-party enrichment. Fair, and stated up front.
The one they do not name: their customer base is teams that chose Linear, which skews toward startups and product-led companies with above-average tooling appetite. Adoption numbers from that population are a leading indicator, not an average. If you read "CEOs at 201+ companies went from 9% to 36%" and picture the median company that size, you will be wrong. The median company that size is not on Linear.
Also worth holding lightly: the time-spent metrics are derived from product telemetry, which measures time in Linear, not time working. A founder spending 26 more minutes commenting could be coordinating more or could be doing coordination in Linear that used to happen in Slack.
What I would actually do
Instrument your own version of this before you trust anyone's benchmark. For four weeks, track two numbers: PRs or shipped units per week, and hours you personally spend reviewing and directing rather than building. If the second grows faster than the first, the agent is not paying for itself no matter what the throughput chart says.
Linear says in its closing note that measuring pull requests indicates motion rather than value, and that future reports will try to reach outcomes. They are right that PRs beat tokens as a metric. They are also right that it is not the finish line. Until someone publishes the outcomes version of this report, the honest position is that we have good evidence AI increases output and no evidence at all about whether that output is worth more.
Author
Lukas
@lukcombinator