By Hamza Faisal
Microsoft Research has open-sourced Agensh, a multi-agent harness that coordinates as many as 1,024 AI coding agents on a single software project — with no central orchestrator telling them what to do. The first version of the system was released this week, turning a September research paper into code any developer can download and run.
Most multi-agent systems today work like a traditional team: one lead agent plans the work, hands out tasks and merges the results. Agensh throws that structure out. Every worker runs the same five-step loop — gather context, claim a sub-task, act and share findings, verify the result, merge progress — and the workers coordinate themselves through shared infrastructure.
No boss, just a shared workspace
The coordination rests on three pieces: a shared workspace holding proposed, ongoing and completed work; a message interface for worker-to-worker communication; and a shared context board that keeps reusable findings and intentions. Because no single agent bottlenecks decisions, the system can scale far beyond what orchestrator-led designs handle — the paper cites Claude Code agent teams, Codex sub-agents, Copilot fleet and Kimi Agent Swarm as examples of the centralized pattern.
The setup runs on Linux with Docker Compose; a single node runs eight workers by default, while the headline 1,024-agent run spread 64 agents across 16 nodes. Each worker gets a 272,000-token input limit, a 128,000-token output limit and a 64 GiB memory ceiling.
More agents, better code — up to a point
To test scaling, the team set Agensh loose on the five hardest tasks in ProgramBench, rebuilding large open-source projects — FFmpeg, gromacs, pandoc, PHP-src and ctags — from scratch under a six-hour, no-internet budget. Scaling from one agent to 128 lifted the mean test pass rate from 19.31% to 28.78%, roughly a 49% relative improvement, and larger teams reached the same scores sooner.
The sharpest result came on pandoc: one agent scored 33.89%, while 1,024 agents reached 55.06% — a 62% relative gain. With a different model, GPT-6.1-sol, just 32 agents hit a 79.59% pandoc pass rate, the best in the study. But the gains taper: going from one agent to eight added only 1.37 points on the five-task mean, a reminder that throwing more agents at a problem is not free — every extra worker costs tokens, compute and coordination overhead.
Agents that organize themselves
Perhaps the most striking finding is what the agents did on their own. As the organization grew, worker trajectories showed cooperation emerging and standardizing in layers: peer coordination first, then multi-worker integration, then standardized workflows, and finally specialized roles — an organization chart that nobody programmed.
The authors frame agent count as a new scaling dimension, most useful for complex tasks under hard time limits. The caveat matters too: these are research benchmarks, not production proof, and a thousand-agent run demonstrates the ceiling, not a recommended default. Still, for teams building coding-agent fleets, Agensh offers a working reference design for what happens when the boss steps aside.
