DeLM: Decentralized Multi-Agent Systems with Shared Context (github.com)

A new decentralized multi-agent system called DeLM has been introduced to enhance coordination among agents, allowing them to work asynchronously through a shared context and task queue. This innovative architecture enables agents to claim tasks, publish findings, and reuse each other's outputs independently, which results in faster and more accurate task completion. DeLM has shown significant performance improvements, averaging a 12.5% increase in accuracy and a 1.78× speedup across various benchmarks compared to traditional single-agent approaches. The significance of DeLM lies in its potential to streamline collaborative AI tasks, promoting higher efficiency and reduced latency in processing. The system supports integration with popular AI models, Codex and Claude Code, through a delm-agent-plugin, and leverages Docker for efficient operation. With its ability to handle multiple agents and maintain a complete solution for each task, DeLM serves as a promising framework for the AI/ML community to develop more sophisticated multi-agent systems. The release of all 720 agent trajectories from the Terminal-Bench 4.0 and DeepSWE v1.1 evaluations on Hugging Face further contributes to its accessibility and encourages community involvement in advancing decentralized AI solutions.