Coding agents shipped 1,400 percent more code on GitHub in 2026 than in all of 2025. That number, disclosed by GitHub COO Kyle Daigle on the Latent Space podcast in a 90-minute interview published June 2, is the first hard infrastructure figure attached to what the industry has been calling the “agentic coding era.” It is not a benchmark. It is a platform throughput count, and it has a concrete consequence: GitHub’s systems were engineered for humans.

Daigle stated the current rate plainly. One billion commits shipped across all of 2025. The current pace is 275 million per week, on track for 14 billion this year if growth holds linear. “It’s still speeding up,” he said when asked whether the April figure was still accurate. GitHub Actions, the platform’s CI/CD compute layer, has grown from 500 million minutes per week in 2023 to 1 billion minutes per week in 2025. The agents are not just writing code. They are triggering builds.

The infrastructure was not ready. Daigle described the problem precisely: the systems underneath GitHub were calibrated for human cadence, where a developer pushes every few minutes and opens a PR every few hours. Agent cadence breaks that model. Bursts of commits arrive in seconds, then stop, then burst again. The permissioning layer, which Daigle said sits in a database internally called MySQL One, was built years ago and has not been fully replaced. That is where the well-publicized outages have occurred. “We’re talking about unique, novel permission problems that happen only at a scale across all different objects,” he said, “that now we have to go rewrite this underlying system.”

The architectural pivot worth understanding is what Daigle describes as the drift away from the fork-and-PR loop. When 80 percent of PRs originate from agents rather than human contributors, the social contract underpinning the review model begins to erode. The pull request was GitHub’s contribution to standardizing distributed software collaboration. It encoded a trust heuristic: a named human with a commit history vouches for a diff. That heuristic does not transfer cleanly when the committer is an agent acting on behalf of a developer. “Trust is kind of diffuse,” Daigle said. “We still tend to use human signals” to decide whether a PR is safe to merge.

The branch-as-task pattern is the architectural response GitHub is moving toward. Instead of a developer forking and submitting a PR for human review in the traditional sequence, the emerging model has an agent claiming a branch, executing against it autonomously, and returning a diff for a single human review gate at the end. The repository becomes the runtime. The branch name is the task descriptor. The diff is the work product. This shifts the bottleneck from developer cycle time, which agents have already collapsed, to review attention and platform throughput, which remain human-speed constraints.

Daigle is candid that the identity problem is not solved. When an agent commits, the provenance question is open: whose intent does that code represent? GitHub is building toward agent accounts linked to human principals, but the social and legal frameworks for trusting agent-authored code in regulated environments lag behind the technical ones. “Even when it is fully verified,” he said, “you probably have to have trust from multinational governments and regulating agencies.”

GitHub’s public outages this spring were the visible symptom of a deeper rebuild. The company is not just scaling horizontally. It is reconceiving what a commit, a branch, and a review mean when the author is not a human sitting at a keyboard. The context briefing notes from the Latent Space interview are accurate on the core claim: the bottleneck has moved. It used to be how fast developers could write code. Now it is whether the platform and the review layer can absorb what agents produce.

Teams currently running high-volume agent workflows on GitHub Actions should treat CPU capacity as a first-order planning constraint for the rest of 2026. The 14-billion-commit pace means compute pressure on shared runners will not ease until GitHub’s infrastructure rebuild catches up with the growth curve.

Latent Space podcast with Kyle Daigle, GitHub COO (latent.space), 2026-06-02.