SwarmCoder: Multi-Agent Collaboration for Complex Software Engineering

ABSTRACT

Single-agent code generation models frequently fail when confronted with repository-level refactoring and complex architectural planning. We propose SwarmCoder, an Agentic AI framework that deploys a specialized hierarchy of agents—including a Planner, a Coder, and a Reviewer—to collaboratively navigate software repositories. By implementing a novel consensus-driven debugging loop, SwarmCoder achieves a 54% pass rate on the SWE-bench-Hard dataset. This paper demonstrates that structured, role-based multi-agent communication is essential for scaling AI to enterprise-level software engineering.