Editorial illustration for Microsoft's Agensh Multi-Agent Harness Scales to 1,024 AI Agents
Microsoft Scales Multi-Agent AI to 1,024 Agents
Microsoft Research has put out Agensh, a harness for running large swarms of coding agents without a boss agent telling them what to do. Most multi-agent setups, including Claude Code agent teams, Codex sub-agents, Copilot fleet and Kimi Agent Swarm, lean on a lead agent that plans tasks, assigns them and stitches the results together. That lead agent becomes the bottleneck once the team gets big enough.
Agensh skips it. Workers pick their own sub-tasks, post their findings for others to see, and push code into a shared Git repo on their own.
The team tested the idea at a scale most labs don't bother publishing: from a single agent up to 1,024, run across 16 nodes with 64 agents each. Each worker gets a 272,000-token input window and a 128,000-token output cap, with a 64 GiB memory ceiling per worker on a Linux/Docker Compose setup. The benchmark is ProgramBench, and the results track what happens to accuracy and speed as the crowd of agents grows, including some sharp swings between best and worst configurations.
Every worker picks its own sub-task, shares findings and merges code into a common Git repo. On ProgramBench, the team scaled the setup from 1 agent to 1,024 agents.
Why this matters
Agensh's real contribution isn't the agent count, it's the architecture choice behind it. Dropping the central orchestrator and letting 1,024 workers claim sub-tasks and merge into a shared Git repo is a bet that coordination overhead, not raw compute, is the bottleneck in multi-agent coding systems. The pandoc numbers back that up: 128 agents hit 30% pass rate in 30 minutes, versus 60 and 90 minutes for 32 and 8 agents.
That's a scaling curve worth watching, not just a benchmark flex. For teams building agent infrastructure, the decentralized design matters more than the headline figure, since orchestrator bottlenecks are exactly what breaks these systems at scale. We'd want to see the GPT-6.1-sol gains (56% to 79.59% on pandoc) replicated outside Microsoft's own harness before treating them as settled.
The absence of these figures from the v1 paper is worth noting too. Open-sourcing the harness means outside teams can now stress-test these claims directly, which is the part actually worth tracking over the next few months.
Related Reading
Browse all Microsoft news →
Research & Benchmarks Microsoft AI Chief: "If It Isn't Safe We Shouldn't Build It
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
Related Topic Claude gains shared context in Excel, PowerPoint; Microsoft adds Copilot Cowork
AI Tools & Apps
Related Topic Microsoft adds ’vibe working’ to Word and Excel; Copilot Agent Mode now default
Business & Startups
From Archives Kimi K3.1 Pro Rivals Claude Opus in Coding and Reasoning
From Archives Study: Fancy coding harnesses cost more but not better across 447 tasks
From Archives Microsoft Proposes Guardian Agents to Secure Overprivileged Enterprise AI
Business & Startups
From Archives Microsoft CEO Nadella warns AI agents are like insider threats
Business & Startups
Common Questions Answered
How does Microsoft's Agensh architecture differ from other multi-agent systems like Claude Code agents and Copilot fleet?
Agensh eliminates the central boss agent that typically orchestrates tasks in multi-agent systems, which becomes a bottleneck as teams scale. Instead, workers autonomously pick their own sub-tasks, share findings, and merge code into a common Git repository without centralized coordination, allowing the system to scale to 1,024 agents efficiently.
What is the main performance advantage of Agensh's decentralized approach on ProgramBench?
Agensh demonstrates superior scaling efficiency compared to traditional hierarchical multi-agent systems, with 128 agents achieving a 30% pass rate in 30 minutes versus 60 and 90 minutes for 32 and 8 agents respectively. This scaling curve shows that removing coordination overhead is more critical to performance than increasing raw compute power in multi-agent coding systems.
Why does Agensh use a shared Git repository for agent coordination?
The shared Git repository enables asynchronous coordination between agents without requiring a central orchestrator to assign and manage tasks. This approach allows each of the 1,024 agents to independently claim sub-tasks and merge their code contributions, reducing coordination overhead and allowing the system to scale effectively.
What does Agensh's architecture reveal about bottlenecks in multi-agent coding systems?
Agensh's design validates that coordination overhead, rather than raw computational capacity, is the primary bottleneck limiting multi-agent system performance. By distributing task selection and code merging across workers instead of funneling everything through a central boss agent, the system achieves dramatically better scaling characteristics.
Further Reading
- Agensh - A Decentralized, Scalable Multiagent Harness - Microsoft Research
- 1,024 Agents, No Orchestrator: The Last 896 Were Worth 4.12 Points - AI Newsway
- CodeScaleBench: Testing Coding Agents on Large Codebases and Multi-Repo Software Engineering Tasks - Sourcegraph
- CooperBench: Benchmarking Cooperation in Coding Agents - ICLR
- SWE-Bench ProMax: Benchmarking Agents on Large-Scale Multilingual Code Refactoring - alphaXiv
