New Dataset Reveals Exactly How GitHub's AI Writes Code — Step by Step
AI is already coding for tech companies. Now we can see how it works.
GitHub's Copilot cloud agent is not a simple autocomplete. It's an AI worker that can take over tasks like exploring a code repository, editing files, running tests, and even opening pull requests. Until now, companies only saw the final results of what this AI produced — not the messy journey it took to get there.
AgentLogs changes that. It's a massive public dataset containing 64 million activity logs from 549,000 AI work sessions across 35,000 popular open-source projects. Every step is recorded: the prompts given, the AI's internal reasoning, every file edit, every command it ran, and even how much computing power it consumed.
Why should you care? AI coding assistants are already being used in real companies, so their failures directly affect software quality, costs, and deadlines. With AgentLogs, engineers can finally study why an AI gets stuck, where it wastes money, and when it produces flawed code. It's like giving managers a security camera view of a "ghost employee" who's been working on the company's products.
The catch: this is a research tool, not a consumer product. The data comes from public repositories, so it doesn't include private company code. And while it shows what the AI did, it doesn't explain every possible glitch. Still, it's the first realistic behind-the-scenes look at how AI works as a digital colleague.
- AgentLogs is the largest public record of AI coding activity ever released, with 64 million logged steps.
- It shows how GitHub's AI explores code, makes edits, and fails — not just what it finishes.
- This helps companies measure the real cost and reliability of AI assistants before trusting them with important projects.
Why It Matters
AI is already building the software you use every day. This research helps make that AI safer, cheaper, and more trustworthy.