AI
New dataset of 1.8M code edits co-authored by AI agents and humans shows promise for training better coding models
Researchers have assembled AgentPack, a corpus of code changes generated collaboratively by AI agents and human developers, demonstrating that AI-assisted edits offer higher-quality training data than traditional commit histories.
A research team has introduced AgentPack, a new dataset comprising 1.8 million code edits drawn from real-world GitHub projects where artificial intelligence agents and human developers worked together. The collection includes contributions from Claude Code, OpenAI Codex, and Cursor Agent through early October 2025.
Why a new approach to training data#
Historically, machine learning models designed for code editing have been trained on commits and pull requests mined directly from public repositories. While this approach seemed logical—commit messages capture human intent in natural language, while code patches show how that intent was implemented—the reality proved messier. Traditional commit data suffers from noise: messages tend to be sparse and unhelpful, individual commits often bundle unrelated changes together, and many contributions come from automated tools following rigid rules rather than reflecting genuine problem-solving.
The advantage of agent-assisted code#
The emergence of modern software engineering agents has shifted this equation. When humans and agents collaborate to produce code changes, the accompanying documentation tends to be far more detailed and explicit about the reasoning behind each modification. Equally important, human gatekeeping provides implicit quality filtering: project maintainers review and selectively reject low-quality pull requests before they land in public code repositories, leaving behind a curated collection of genuine improvements.
Performance gains over prior training corpora#
The researchers demonstrate that language models fine-tuned on AgentPack achieve better performance than models trained exclusively on human-authored commits from earlier datasets. This finding suggests that the quality and clarity of agent-assisted edits—particularly the richer natural-language descriptions that accompany them—translate into more effective training material for building future code-generation systems.
The work charts broader trends in software engineering agent adoption and examines the structural characteristics of the edits, providing a foundation for understanding how human-AI collaboration is reshaping the landscape of code development.
