Show HN: Agentsh – A Decentralized, Scalable Multiagent Harness (github.com)

Agentsh, an innovative decentralized multi-agent harness, has been open-sourced, allowing researchers and developers to explore its capabilities in scalable agent collaboration. Capable of accommodating up to 1,024 agents, Agentsh enables self-organized peer coordination, where agents autonomously claim sub-tasks, share insights, verify their findings, and merge progress without a central orchestrator. The system's architecture—a shared workspace combined with dynamic messaging and context-sharing—facilitates efficient asynchronous interactions among agents, crucial for complex problem-solving. The significance of Agentsh lies in its demonstrated ability to enhance performance on challenging tasks, such as those in the ProgramBench benchmark. When evaluated using advanced models like GPT-5.6-sol and GPT-6.1-sol, Agentsh achieved notable improvements in task success rates, showcasing a nearly 50% increase in efficiency with larger teams. This self-organizing framework could revolutionize multi-agent systems by enabling scalable and effective collaboration, potentially impacting areas such as distributed AI, robotics, and automated programming. The release includes comprehensive resources for replication and integration, underscoring the potential for widespread adoption and innovation in the AI/ML community.