GitHub's Octoverse: 40 percent of pull requests now start with an agent, and 61 percent of Copilot users switch models mid-session
AspireAI community · GitHub
GitHub's annual state-of-open-source report lands with the clearest numbers yet for how far agentic coding has moved into everyday work. Developers pushed code 17.9 billion times in the past 12 months, more than 2024 and 2025 combined, and commits across public and private repositories climbed to 34.3 billion. Merged pull requests more than tripled to 1.3 billion, an average of 3.5 million a day, and Actions ran a job 1,440 times every second.
The agent numbers are the headline. By September, more than 40 percent of pull requests started with an agent's first draft, up from 17 percent in March, and agent-authored pull requests merged within two weeks at 90 percent against about 86 percent for the comparison group. Repos created this year with an agent configuration file logged eight times as many commits as the average new repo without one. Among Copilot's monthly active users, 55 percent now use agentic features, up from 35 percent a year ago, and the hands-off mode of assigning a task to an agent grew 230 percent year over year against 85 percent for agent chat. On model choice, 61 percent of users who picked a model themselves used more than one in September, and no single model carried more than 25 percent of requests.
Two honest caveats. GitHub measures its own platform, so the counts describe GitHub users rather than software development in general, and the agent-detection method is GitHub's own. The report also notes that its own forecasting models no longer agree on where developer growth goes next: across more than 40 models run forward to 2031, about half call for extreme growth of more than three times and only three expect anything below double.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.