SWE-agent-2.0: Benchmarking Multi-Agent Swarms on Full-Stack Software Engineering

ABSTRACT

As single-agent coding assistants plateau in their ability to handle repository-scale refactoring, multi-agent swarms have emerged as the new standard for autonomous software engineering. We introduce SWE-agent-2.0, a framework and evaluation suite designed to benchmark distributed agent swarms on real-world GitHub issues. By assigning distinct roles (Architect, Developer, Tester) within a sandboxed Unix environment, our reference swarm achieves a 42% resolution rate on the SWE-bench-Lite dataset, demonstrating that specialized agentic collaboration significantly outperforms monolithic code generation.