EvoGit: Decentralized Code Evolution via Git-Based Multi-Agent Collaboration

We introduce EvoGit, a decentralized multi-agent framework for collaborative software development driven by autonomous code evolution. EvoGit deploys a population of independent coding agents, each proposing edits to a shared code-base without centralized coordination, explicit message passing, or shared memory. Instead, all coordination emerges through a Git-based phylogenetic graph that tracks the full version lineage and enables agents to asynchronously read from and write to the evolving code repository. This graph-based structure supports fine-grained branching, implicit concurrency, and scalable agent interaction while preserving a consistent historical record. Human involvement is minimal but strategic: users define high-level goals, periodically review the graph, and provide lightweight feedback to promote promising directions or prune unproductive ones. Experiments demonstrate EvoGit's ability to autonomously produce functional and modular software artifacts across two real-world tasks: (1) building a web application from scratch using modern frameworks, and (2) constructing a meta-level system that evolves its own language-model-guided solver for the bin-packing optimization problem. Our results underscore EvoGit's potential to establish a new paradigm for decentralized, automated, and continual software development. Large language models (LLMs) have significantly advanced the automation of software development, excelling at tasks such as code generation, debugging, and documentation Zhao et al. (2023); Minaee et al. (2024); Y ang et al. (2024a); Team et al. (2023); OpenAI et al. (2024). Building upon this foundation, recent efforts have embedded LLMs into autonomous coding agents Qian et al. (2023); Y ang et al. (2024b); Hong et al. (2024); Islam et al. (2024), enabling the execution of multi-step development workflows through tool usage, memory, and interaction policies Y ao et al. (2023); Xi et al. (2023); Wang et al. (2024). While promising, current frameworks face critical limitations. Most rely on scalar reward signals, ground-truth unit tests, or dense human feedback to supervise agent behavior. These assumptions are ill-suited for open-ended software engineering, where success criteria are emergent and multifaceted. Furthermore, agent collaboration is typically centralized and synchronous, requiring a global orchestrator and shared memory, which restricts scalability and robustness. Critically, the development process often lacks traceability: as code branches diverge and recombine, it becomes increasingly difficult to understand the provenance of design decisions or to reproduce successful outcomes.