Source note

Investigating Autonomous Agent Contributions in the Wild: Activity Patterns and Code Change over Time

Autonomous Coding AgentsSoftware EngineeringGithub Pull RequestsCode ChurnHuman AI Collaboration

This paper studies how autonomous coding agents contribute to real GitHub projects and how their code changes after merge. Using a new dataset of about 112k pull requests across five agents and a human comparison set, it finds that agent use is growing, especially in low-star repositories, and that agent-authored code shows higher later churn than human-authored code.

  • Existing evaluations of coding agents rely heavily on benchmarks, user studies, or small datasets, so they miss how agents behave in real collaborative software work.
  • Prior field studies often cover one agent, a small PR sample, or only popular repositories, which makes it hard to compare agent behavior at scale.
  • Code generation is only part of software engineering; maintainability matters because code must survive and be edited over time after the pull request is merged.
  • The authors build a GitHub dataset of 111,969 pull requests from June-August 2025, covering five autonomous agents: OpenAI Codex, Claude Code, GitHub Copilot, Google Jules, and Devin, plus a matched human-authored PR set.
  • They identify agent PRs with concrete GitHub signals such as branch prefixes (head:codex/, head:copilot/), bot authors (google-labs-jules[bot], devin-ai-integration[bot]), and Claude watermark text in PR descriptions.
  • The dataset includes PRs plus linked commits, comments, reviews, issues, and changed files, with counts in the tens of thousands per agent and millions of lines of code overall.
  • They compare agent and human PRs on collaboration and activity measures such as merge frequency, merge latency, edited file types, change size, commit density, comments, reviews, and repository characteristics.
  • They also run a longitudinal analysis of post-merge code evolution using survival and churn estimates to measure how much agent-authored code is later kept or rewritten.
  • The final dataset contains 111,969 PRs: 20,835 Codex, 19,148 Claude Code, 18,563 Copilot, 18,468 Jules, 14,045 Devin, and 20,910 human PRs.
  • Associated activity volume is large: for example, the dataset includes 102,037 human commits, 82,755 Claude commits, 69,896 Copilot commits, 51,641 Devin commits, 41,032 Jules commits, and 27,530 Codex commits; changed files range from 90,822 for Codex to 255,275 for Claude.
  • The paper claims agent activity in open-source projects is increasing and is concentrated more in low-star repositories than earlier studies suggested.
  • The main substantive finding is that agent-authored contributions are associated with more code churn over time than human-authored code, which points to lower long-term stability or more follow-up maintenance.
  • The excerpt does not provide the actual churn, survival, merge-rate, or latency numbers, so the quantitative size of those differences cannot be verified from the provided text.
  • As a concrete research asset, the authors release the dataset publicly on Hugging Face for follow-up work on agentic software development.