The useful signal this morning is not that coding agents can produce more code. It is that the work around that code has not disappeared.
The volume jump is large. Linear's first How Teams Build report uses activity from paid workspaces to follow software work from an issue to the pull request that closes it. In a fixed group of 6,887 teams, those with a coding agent connected went from an average of 21 pull requests—proposed code changes—opened per week in June 2024 to 65 in June 2026. Teams without one moved from 8 to 10. Across all paid workspaces, weekly pull requests per team rose 111% over the same two years.
The change begins before code. During the week of August 3, agents and other automated clients created 2.435 million work tickets in Linear, just short of the 2.481 million created by people and ordinary integrations. Two years earlier, machines created fewer than one issue in a thousand.
Linear is unusually clear about the boundaries of those numbers. Agent-connected teams were already higher-output before coding agents existed, so this is a correlation, not an experiment. The report counts pull requests opened, not merged, and sees only repositories connected to Linear. An opened pull request can be useful, trivial, rejected, or harmful. It measures motion, not value.
The surrounding work did not get cheaper. Between June 2025 and June 2026, average engineering time inside Linear spent creating and triaging issues rose from 24 to 28 minutes per user per month. Commenting rose from 35 to 40 minutes. Measured planning time stayed roughly flat, while chatting with AI and delegating issues appeared as a new category on top.
These are small slices of a developer's month inside one product, not total working hours, and the comparison does not prove agents caused the change. The defensible conclusion is narrower: the visible coordination work did not contract while software activity accelerated. More output brought more material to specify, explain, sort, and review.
Another dataset points in the same direction. A July study of Microsoft's early rollout of Claude Code and Copilot CLI found that adopters merged about 24% more pull requests than a modeled counterfactual over four months. That is stronger than counting opened pull requests, but still not a quality measure. The authors note that merged-PR counts reward small changes and can miss complexity and defects; the study covers one company, and all three authors work at Microsoft, which owns GitHub and sells Copilot.
The scale makes the unanswered part urgent. JetBrains' separate survey of more than 15,000 professional developers reports that 68% now use AI coding agents daily. That is self-reported vendor research, and its “coding agent” category is broad, but it makes this more than a question about a few early adopters.
The bottleneck moved. When a machine can print more pages, an editor's attention becomes more valuable, not less. Software teams now need to measure the parts that raw output hides: how many changes merge, how long review takes, how often code is reverted, whether defects rise, how much human time remains, and what the models cost. If pull requests double while the review queue and failure rate grow with them, the team is busier, not necessarily more productive.
What to watch. Linear says later editions will connect token spend to more of the development lifecycle, including code review. The strong claim today is that coding agents are a volume multiplier. The claim that they reduce work or create proportional value is still unproven.
Source graph: Semble source collection