GitHub says agents now start more than 40% of pull requests—activity is not productivity

GitHub’s 2026 Octoverse report contains the clearest platform-scale evidence yet that coding agents have moved into ordinary software work. By September, more than 40% of pull requests on GitHub began with an agent’s first draft, up from 17% in March. That amounted to 110 million agent-started pull requests in September alone, roughly six times the March volume.[1]

The platform grew at the same time

GitHub reports that the average repository created in 2026 with an agent-configuration file recorded eight times as many commits as a new repository without one. Agent-started pull requests were also more likely to change multiple files—87% versus 79%—and by September 90% were merged within two weeks, compared with about 86% for pull requests without detected agent involvement.[1]

Those figures describe a substantial workflow shift. They do not identify its cause. A repository with an agent configuration may belong to a larger, more active team; a pull request can be generated, revised or split automatically; a higher commit count can reflect finer-grained changes rather than more delivered value. Fast merging can indicate useful work, lighter review, smaller risk or simply different task selection.[1]

“Agent-started” is a useful but incomplete label

GitHub defines an agent-started pull request as one whose first draft came from an agent the platform can detect. The report calls its agent signal a floor because not every use is visible. The label also does not mean an agent independently chose the task, designed the change and approved the result. A human may have written a detailed specification, reviewed every line and requested several corrections.[1]

That is why the report is stronger evidence of orchestration than autonomy. Developers are increasingly beginning work by describing the desired outcome and then shaping the returned change through review. GitHub says 55% of Copilot monthly active users now use agentic features, while its cloud agent grew 230% year on year.[1]

Model choice is becoming part of development practice

In September, 61% of Copilot users who selected models used more than one. Among developer-selected requests, GitHub says 38% went to general-purpose models, 37% to deep-reasoning and planning models, and 25% to fast lightweight models. Separately, 63% of paying Copilot developers used automatic model selection at least once.[1]

This suggests that developers are not converging on one universal coding model. They are trading capability, latency and cost across planning, implementation and review. That makes the agent harness—and its ability to route work, preserve context and expose diffs—as important as benchmark rankings of the underlying model.[1]

The missing denominator is delivered software quality

Octoverse is based on GitHub telemetry across public and private repositories, with different public-only analyses for industries and open source. It measures events that GitHub can observe. It does not measure whether a feature satisfied users, reduced defects, improved maintainability or saved a developer an hour after review costs are counted.[1]

The appropriate conclusion is therefore narrower than “agents made developers eight times more productive”. Agents are involved in a rapidly growing share of repository activity, and teams using them exhibit different work patterns. Establishing productivity requires controlled comparisons or careful longitudinal studies that include rework, defects, security, review time and outcomes—not only commits and pull requests.[1]

Even with that limitation, the report marks an inflection. Software production on the world’s largest code platform is no longer adequately described as humans writing code with occasional autocomplete. A growing share begins with delegated work and ends with human judgement about whether that work belongs.[1]

Report an error or suggest a correction