AI Coding Agents have more than 2m of PRs on GitHub with 80%+ acceptance rate (prarena.ai)

AI coding agents have already authored more than 2 million GitHub pull requests and—when measured by “ready” PRs—show over an 80% merge/acceptance rate. The way agents work differs substantially: some (e.g., Codex) iterate privately and submit mostly ready-to-review PRs, producing few drafts but high merge rates, while others (e.g., Copilot, Codegen) create draft PRs to enable public iteration before marking them ready. To enable a fair apples-to-apples comparison across these workflows, the default success metric reports acceptance using Ready PRs only; you can toggle “Include draft PRs” to reveal the full volume and revision history of agent activity. This distinction matters for the AI/ML community because evaluation and trust depend on how success is measured. Ready-PR metrics emphasize an agent’s ability to produce mergeable code, but they can undercount collaborative or incremental development workflows that surface drafts and feedback. Conversely, including drafts gives a fuller picture of agent-assisted development velocity, reviewer burden, and iteration patterns. For researchers and tool builders, the takeaway is to report both ready-only and draft-inclusive metrics: they reveal different strengths (final code quality vs. collaborative, iterative assistance) and have implications for benchmarking, repository hygiene, and automated code review pipelines.