THE 60-SECOND VERSION
● AI writes just under half of everything created in Linear. Two years ago it was fewer than 1 in 1,000.
● The finding nobody expected: total product development time still went up. Agents tripled pull requests and added work rather than replacing it.
● Pennsylvania made local approval legally binding for AI data centres and banned NDAs on the projects.
● 10 days left on Claude Sonnet 5 at 2 dollars per million input.
The productivity paradox, measured
Linear published a report drawn from aggregated product data across its paid workspaces — AI conversations, agent sessions, issue activity, comments and pull requests. It compares June 2025 against June 2026.
The headline number is startling on its own. Two years ago fewer than one issue in a thousand was created by AI. Teams now use AI to write just under half of everything created in Linear, and on current trajectory it will shortly author more than people and integrations combined.
| Metric |
Change |
| Issues authored by AI |
Under 0.1% to just under 50% |
| Weekly PRs, teams using agents |
21 to 65 |
| Weekly PRs, teams not using agents |
8 to 10 |
| Engineering time on create and triage |
Up roughly 17% |
| CEO adoption, companies over 201 people |
9% to 36% |
| Total product development time |
Increased |
THE LINE THAT MATTERS
A team tripling its pull request throughput while total development time rises is not a team going faster. It is a team doing more things. The agents added work rather than removing it, and the coordination overhead of directing them landed on the same engineers.
Time spent creating, triaging and commenting rose in nearly every function. Founders swung hardest — up 17 minutes on issue creation and 26 on commenting, though that cohort is small and noisy. Linear frames it as more work needing more coordination, with that coordination increasingly setting the context agents act on.
We dug into what the report actually implies for teams running agents, and what to change, separately.
Pennsylvania puts data centres under local control
Governor Josh Shapiro signed Executive Order 2026-05 on 18 August, making the state's GRID standards legally binding on data-centre developers.
Projects now require a Consent Order committing the developer to local approval, full funding of any new electricity infrastructure, water-conservation measures and local hiring. Every AI data-centre proposal has been pulled from the state's Fast Track permitting programme, and NDAs on data-centre projects are prohibited.
The reversal is the notable part. Shapiro had previously championed a 20 billion dollar Amazon buildout and the fast-track process that enabled it. The NDA ban is the clause with the longest reach — much of the local opposition to these projects has been driven by residents discovering terms only after approval.
Cerebras claims 30x on inference
Cerebras unveiled the CS-4, a rack-scale system on its new Nexus architecture with three WSE-3 Turbo wafers per rack, each roughly doubling the previous generation. The company claims up to 30 times faster inference than GPU systems and more than 1,000 tokens per second on 10-trillion-parameter models. First shipments begin this quarter.
Treat the 30x as a vendor figure until someone independent runs it. What is worth watching is the 10T-parameter reference point — that is a scale above anything publicly shipped, which tells you what Cerebras expects to be serving.
Also worth knowing
- Stripe engineers described Minions, autonomous coding agents generating over 1,300 pull requests per week, with tasks originating from Slack, bug reports or feature requests.
- Spotify presented Honk at QCon London, an agent handling code migrations across its codebase, cutting migration timelines on work traditional scripts could not handle.
Deadlines still running
| Date |
What happens |
| Aug 31 |
Claude Sonnet 5 moves 2 to 3 dollars per million input, plus a tokenizer change adding 10 to 35 percent tokens on code |
| Aug 31 |
kimi-k2.5 and moonshot-v1 sunset, migrate to kimi-k3 |
| Early-mid Sept |
Grok 4.7 window per Musk. No model ID at docs.x.ai yet |
| Oct 1 |
OpenAI vs Apple hearing |
| Oct 24 |
deepseek-chat and deepseek-reasoner deprecated |
What we are watching next
Whether anyone publishes the counter-example. Linear's data covers its own paid workspaces, which skews toward software teams already inclined to adopt tooling. If a team somewhere has tripled throughput and cut cycle time, that report has not surfaced — and its absence is doing a lot of quiet work in this debate.
FAQ
Does AI really write half of all Linear issues?
Just under half of everything created, per Linear's aggregated data across paid workspaces, comparing June 2025 to June 2026. The figure covers agents and MCP integrations against people and standard integrations.
If agents triple pull requests, why did development time rise?
Because throughput and cycle time are different measurements. More PRs means more review, more coordination and more context-setting, and that work landed on the same engineers. Linear reports time spent creating, triaging and commenting rose in nearly every function.
Is the Linear data representative?
Partly. It covers paid Linear workspaces only, which skews toward software teams that already adopt tooling early. The direction is credible; the magnitude may not generalise to teams outside that profile.
What did Pennsylvania actually change?
Executive Order 2026-05 makes GRID standards legally binding, requires a Consent Order covering local approval, electricity infrastructure funding, water conservation and local hiring, removes AI data centres from Fast Track permitting, and prohibits NDAs on these projects.
Should I believe the Cerebras 30x claim?
Not until it is independently measured. It is a vendor benchmark against unnamed GPU systems. The specific figure worth noting is 1,000 tokens per second on 10-trillion-parameter models, which implies a serving scale above anything publicly shipped today.