Agents Outpaced Humans in GitHub PRs and Linear Tasks in August 2026

In August 2026, agents created more pull requests on GitHub than humans did. In Linear, agents create more tasks than live employees do. Gergely Orosz, author of The Pragmatic Engineer newsletter, presented these figures in a talk at the LDX3 conference in New York. He also described how tools are changing: Cursor and Antigravity are moving away from the classic IDE, and in many companies code review has turned into theater.

We have seen the original talk only through a recap, so what follows covers the points Orosz puts front and center. For detailed figures and specifics, check his full text.

Humans are no longer the main authors of code

According to Orosz, the turning point came in the summer of 2026. On GitHub, agents now open more PRs than developers do. In the Linear tracker, they also create more tasks than people do. The agent is no longer an assistant that completes a line in the editor. It defines the work, does the work and packages the result on its own.

To support this, Orosz uses fresh data from GitHub, Linear and Factory AI. Stories from OpenAI, Anthropic, Uber and Ramp add to it: companies where agentic workflows are already built into everyday development.

The IDE loses its central place

The second shift concerns the interface. Cursor and Antigravity, two prominent AI development tools, have moved away from the familiar "editor with a chat on the side" model. If agents do most of the work, a human needs not a text editor but a panel where they assign tasks, review results and make decisions.

This also changes the economics of attention. A developer writes fewer lines but makes more decisions per unit of time. The more agents run in parallel, the more pressing the question of who checks their output, and how.

Code review has turned into theater

The most worrying point of the talk concerns code review. When PRs arrive in a stream and the reviewer has neither the time nor the context, review becomes a ritual: open the diff, scroll through it, click Approve. The process is formally followed, but in practice no one guarantees quality.

The problem is scale. An agent creates a PR in minutes, while a careful review takes as long as it used to. The gap between the speed of generation and the speed of review keeps widening.

What the recap tells us

• August 2026: agents on GitHub created more PRs than humans.
• Linear: agents create more tasks than humans.
• Cursor and Antigravity are moving away from the IDE model.
• In many companies, code review has become a formality.
• Data sources: GitHub, Linear, Factory AI.
• Case studies: OpenAI, Anthropic, Uber, Ramp.

Practical takeaways for teams

These are our editorial recommendations, not quotes from the talk. First, measure the share of PRs and tasks that come from agents in your team. Without that number, you will not know where the bottleneck really is. Second, rethink review. Automated checks, tests and static analysis should filter out the obvious, so that a human looks at architecture and risks rather than style. Third, limit the flow. If reviewers cannot keep up, a cap on the number of agent PRs per day is better than meaningless Approves. Fourth, evaluate tools by how well they help you manage agents, not by how comfortable the editor is.

The main question of 2026: who on your team is responsible for the result when an agent writes the code and review has become a formality.