Agensh enables scalable multi-agent coding without a lead agent
Summary
A new Microsoft paper has revealed that coding agents can perform more efficiently without a central manager, as demonstrated by the Agensh system. When tasked with the five most difficult challenges from ProgramBench, the study found that increasing the number of self-organizing agents from 1 to 128 improved the average task completion score significantly, showcasing a 49% relative improvement. By enabling agents to claim and execute their own sub-tasks independently while coordinating through shared tools like a Git repository, Agensh offers a substantial scalability advantage over traditional multi-agent systems that depend on a lead agent for task management.
Analysis
Agensh: Agensh is a scalable self-organized multi-agent harness for AI coding agents that operates without a central orchestrator. It enables concurrent workers to gather context, claim and self-assign sub-tasks, execute actions, share findings, verify results, and merge progress asynchronously using shared workspace, messaging, and context components. The Microsoft paper introduces Agensh to overcome scalability limits of prior harnesses that rely on a single lead agent for task allocation and coordination. Task Coordination: Agents coordinate through shared tools including a Git repository for merging work and a shared board for logging findings, supporting fully decentralized execution. Scalability Approach: Self-organized multi-agent cooperation allows performance gains on complex tasks by scaling the number of agents rather than relying on centralized management.
Categories
ai_agentsmachine_learningtech