Back to Feed
Agents / Benchmarks & Evals

How GitHub Projects Adopt Agentic Coding

Original: Early Adoption of Agentic Coding Tools by GitHub Projects
By Maliha Noushin Raida, Daqing Hou Published 2026-07-15

Listen to the summary

Uses a voice available on your device

Audio options
On this page 4 sections
Related concepts 1 concepts

Key Takeaways

  • Most repositories show low agentic activity, with a median of only one to two agentic pull requests every three months.
  • Only 1% of the studied projects exceeded the industry baseline of 36 pull requests per human participant over a three-month period.
  • The dominant mode of operation is single-human oversight, representing 78.9% of all agentic pull requests where one developer handles both review and commit tasks.
  • The research utilized the AIDev-pop dataset to categorize human-agent collaboration patterns across projects of varying sizes.

Summary & Methodology Analysis

The researchers investigated agentic tool adoption by analyzing pull requests from 2,361 repositories within the AIDev-pop dataset, ensuring each project had at least 100 stars. They stratified these projects into three groups based on contributor count: small (1 to 5), medium (6 to 15), and large (16 or more). The team identified agentic contributions by tracking interactions with tools like GitHub Copilot, OpenAI Codex, and Claude Code, while mapping human roles such as reviewers and committers through commit metadata and API logs. This allowed them to classify human-agent collaboration into five distinct patterns based on the division of labor between human actors and AI agents. The study normalized productivity by calculating the volume of pull requests per human participant over a three-month window. The analysis revealed that while agentic tools are integrated into various project sizes, the overwhelming majority of activity is characterized by a solitary developer managing the entire lifecycle of an agent-assisted change. Specifically, 78.9% of agentic pull requests involved a single human responsible for both review and commit, suggesting that these tools currently function as individual productivity aids rather than team-based collaborative systems. A significant limitation is that the research does not measure the qualitative complexity of the code produced or the specific effort required for review, as the analysis focuses purely on PR volume. Furthermore, the findings are restricted to repositories with over 100 stars and may not generalize to smaller open source projects or private, enterprise-grade development environments.

Interactive System Flowchart

Click diagram to expand and zoom

Cross-Examination & FAQs

A deeper dive clarifying mechanics, constraints, and baseline evaluations.

Q1. What is the primary focus of this research?

The paper studies how developer teams adopt and manage agentic coding tools in terms of productivity and collaboration.

Q2. Which tools were included in the scope of the study?

The study analyzed repositories using GitHub Copilot, OpenAI Codex, and Claude Code.

Q3. What is the general trend in tool adoption found by the authors?

Most projects demonstrate low usage, with a median of only one to two agentic pull requests per three-month period.

Q4. How did the researchers categorize repositories for their analysis?

Repositories were stratified by contributor count into small (1-5), medium (6-15), and large (16+) groups.

Q5. What is the industry-reported baseline for agentic productivity?

The baseline cited in the paper is 36 pull requests per participant over a three-month period.

Q6. Did the study find widespread high-volume adoption of these tools?

No, only 25 out of the 2,361 projects studied, which is 1%, exceeded the industry-reported baseline.

Q7. What is the most common pattern of human-agent collaboration?

Single-human oversight, where one developer performs both the review and the commit, accounts for 78.9% of agentic pull requests.

Q8. Does the paper evaluate the quality of the agent-generated code?

No, the paper notes that the productivity analysis does not account for qualitative differences like code complexity or the effort required for reviews.

Q9. Are the results applicable to private repositories?

The authors state that their findings are based on a dataset of repositories with over 100 stars and may not represent private environments.

Browse the archive →
Agents / Benchmarks & Evals Benchmarking AI Agents for Telecom Troubleshooting The researchers introduced CTBench, a new evaluation framework designed to measure how effectively AI agents diagnose and resolve issues in complex, heterogeneous telecommunications network environments. Benchmarks & Evals / Agents A New Benchmark for Detecting Vulnerable Code VICBench is a new multi-language benchmark containing 100 verified vulnerability-inducing commits that helps evaluate automated security detection tools. Benchmarks & Evals / Agents Curated RAG Systems Outperform Frontier LLMs Researchers developed VITA, a domain-specific retrieval-augmented generation system that achieves superior performance on clinical tasks by prioritizing curated local data over generic model scale.