AI Writes Half Your Tickets Now. It Didn't Save Anyone a Minute.

In July 2024, AI agents wrote one issue in Linear. Not one percent. One issue, out of 621 created that week.

Last week, agents wrote 2,435 issues. Humans and integrations wrote 2,481.

That's the crossover. Within a couple of years, the majority of tickets in one of the most popular project trackers on earth will be written by machines — for machines and humans to work together. And buried in the same report is a finding nobody putting out AI ads wants you to see: all this AI made people work more, not less.

Linear just published its first "How teams build" report — six years of first-party telemetry from tens of thousands of software teams, covering adoption, time use, and output from before AI was a thing to right now. No survey. No vibes. Actual behavior, measured. It's the most honest data on AI-assisted development you'll read this year, and almost nobody is talking about the parts that matter.

One issue in a thousand, to half of everything

First, the shape of the curve, because it's absurd.

Two years ago, fewer than one issue in a thousand was created by an AI agent or MCP client. Through late 2024, the weekly count of agent-written issues literally reads: 0, 1, 1, 0, 1, 1, 0. Then 2025 happens. By December, agents are writing 120–147 issues a week. By February 2026, over 500 a week. By July, 2,000+. The last data point — early August 2026 — has agents at 2,435 against 2,481 for everyone else.

Think about what an issue in a tracker actually is. It's the unit of "work that needs doing." It's how a team describes reality to itself. When half of those are authored by agents, your project tracker stops being a shared to-do list and becomes an API where machine coworkers file their intentions. The tracker is turning into the coordination layer between humans and agents — which, conveniently, is exactly why Linear is investing so hard here. But it's bigger than Linear. This is the clearest evidence yet that "AI teammate" stopped being a conference talk and became a Tuesday.

And the humans aren't writing fewer tickets, by the way. Human-and-integration issue creation also grew, from ~600/week in mid-2024 to ~2,400/week now. AI didn't replace the ticket-writing. It piled on.

Everyone adopted it. Even the CEOs.

The adoption data covers 127,000 paid users active in both January and June 2026. In six months, the share of people actively using AI features more than doubled in every single function:

• Product: 12% → 34% (the fastest climbers)
• Engineering: 12% → 30%
• Founders: 14% → 30%
• Design: 6% → 22%
• Go-to-market — the people furthest from a codebase: 5% → 18%

Two details in there deserve more attention than they're getting.

First, company size barely mattered. Adoption roughly tripled at 8-person startups and 10,000-person enterprises alike (8% → 23% and 8% → 25%). That almost never happens with new technology. Usually enterprises lag startups by years on anything with a learning curve. AI is spreading like infrastructure — email, Slack — not like a tool you evaluate and pilot and maybe roll out. Nobody runs a pilot on email.

Second, the executives. CEOs at companies with 201+ employees went from 9% to 36% active AI usage in six months — the single largest jump anywhere in the report, beating every IC function. When a CEO personally starts using a technology instead of reading a McKinsey deck about it, the "is this real" phase is over. You don't learn a tool hands-on because it's hyped. You learn it because it's starting to touch your job.

The output side: agents tripled throughput

Now the part everyone screenshots. Pull requests per team are up 111% over two years against a June 2024 baseline. Flat for the first year, then the curve bends hard upward through 2026 as model quality and adoption climbed together.

But the sharpest cut in the whole report is the cohort comparison: teams that connected a coding agent versus teams that didn't.

• Coding-agent teams: 21 PRs/week in June 2024 → 65 PRs/week in June 2026. Roughly tripled.
• Teams without: 8 → 10.

Linear is careful here, and you should be too — the agent teams were already shipping nearly 3x more before agents existed. This is self-selection: fast teams adopt accelerators. You can't read this as "connect Claude Code, get 3x." But each cohort measured against its own baseline still tells a clean story, and nearly all the growth in the entire dataset sits on the agent side. The teams that were already moving got the boost. The teams standing still mostly kept standing still.

If you're wondering which side of that divergence you want to be on — well. So is everyone else.

The uncomfortable finding: nobody got time back

Here's where the report quietly detonates the marketing narrative.

The entire pitch for AI productivity tools has been time savings. "Give the boring work to the model." "Ship faster, work less." So Linear measured time spent per user on every category of work — creating and triaging issues, assigning and updating, commenting — June 2025 versus June 2026.

Time went up. Almost everywhere.

Engineers spend 5 more minutes a month creating and triaging, 5 more commenting. Founders — the noisiest cohort — spend 17 more minutes creating, 26 more minutes commenting, per month. Design, GTM: up across the board. And here's the kicker from the report: nothing shrank to make room. AI usage appeared as a new layer of work on top of the old work. Total time on product development is going up, not down.

Linear's own head of data calls it what it is: a Jevons paradox. When efficiency increases, consumption of the resource increases too. Better engines didn't reduce coal burning; they made coal cheap enough to burn everywhere. Better code generation didn't shorten the workweek; it made shipping cheap enough that teams ship (and review, and triage, and discuss) far more than before.

If a vendor promised you a four-day week, the measured reality so far is that you got a fourth channel of productivity instead. Whether that's a scam or the whole point depends on what you do with the throughput — but don't let anyone tell you the data shows people working less. It shows the opposite.

Planning didn't move. That's the tell.

My favorite finding is the one nothing happened with.

Minutes spent on customer requests, docs, and long-form planning: essentially flat across every function. Plus one minute here, zero minutes there. In a year when nearly everything else in this dataset moved, the part where humans decide what to build didn't budge.

AI has changed how teams execute. It has not changed how teams decide. The choosing is still ours — and as agents take over more of the doing, the choosing is also where the remaining leverage concentrates. The engineers whose roles evolve fine are the ones whose taste, judgment, and context translate into better instructions for machines. The ones who struggle are the ones whose value was purely in the doing.

That's also, if you're building tooling (hi), the roadmap: everything around execution is being commoditized by agents right now. Everything around deciding — context gathering, prioritization, taste — is wide open.

The people who used to describe changes now ship them

One more role-blurring data point: non-engineers are attaching pull requests at 3x the rate of two years ago. PMs went from 3% to 10%. Designers from 1% to 8%. Linear's phrasing is perfect: "The people who used to describe a change increasingly ship it themselves."

Take your pick of interpretations. It's empowerment — the PM who used to write a spec, wait three weeks, and get something vaguely related back can now just... make the thing. Or it's the death of specialization, scope creep wearing a smile, with product managers doing quiet unpaid engineering at 11pm. Honestly, it's both, and the report's own conclusion leans into it: "everyone in an organization is becoming a builder" seems to be directionally true.

Note what that does to the org chart, though. If PMs ship code and CEOs do IC work and agents write half the tickets, then "engineer" stops being a location in the org and becomes a skill lots of people have some of. That transition is either going to be smooth or extremely not, depending on the company.

Read the fine print before you quote this

Credit where due — Linear is unusually honest about what this data can't tell you, and you should steal their caveats before citing the numbers:

They count PRs opened, not merged. An opened PR says nothing about the value of the change. The report itself concedes that looking at pull requests "indicates motion rather than value" — it's just a much better proxy than token spend, which was the previous best (worst) metric the industry had. A mechanical refactor burns millions of tokens; a one-line fix that saves the company can burn forty.

They only see what happens inside Linear (plus connected repos). Anything outside is invisible, which means every number here is a floor, not a ceiling. And "AI-active" is defined generously — one AI interaction in 28 days counts. Someone who asked Slack bot one question in June is in the 34%.

Also, and this is the part I respect most: Linear explicitly says they have no way of knowing whether the 111% output increase led to positive business outcomes. Acceleration is not the same as progress. You can triple your speed in the wrong direction.

But the direction of travel isn't really in dispute anymore. This is the first serious attempt to measure the AI era with workflow data instead of token counters, and even with all the caveats, the story it tells is coherent: adoption is universal and fast, agents compound for teams already moving, output is way up, and time saved is nowhere to be found.

What I'd take from it

Three things, if you work in software:

The adoption debate is over. It finished while everyone was arguing about it. Your CEO is using the tools. Your PM is shipping PRs. The question isn't whether AI touches your workflow anymore — it's whether you're the one directing the agents or the one competing with them.

Agent leverage is real but it favors the already-fast. The 21-to-65 teams weren't saved by AI; they were amplified by it. If your team is slow for reasons AI can't fix — unclear priorities, bad process, no ownership — agents will just help you produce more of the wrong thing, faster.

And nobody is working less, so stop waiting for the productivity dividend in the form of free time. It arrives as throughput, and you have to go claim it. The four-day week doesn't fall out of the box. The box gives you a bigger engine. Where you drive it is still, stubbornly, a human decision — the one part of the workflow the data says hasn't moved an inch.

If you want to feel what "directing an agent" is like on your own machine — one that browses, runs terminal commands, reads your files, and automates the boring parts of your day, running on open-source models — CopperRiver does exactly that, on your Mac, and starts at $9/month. The teams in that 65-PR cohort connected agents to their workflow. This is the same move, one desk down.