Microsoft Research open-sourced Agensh on October 24 — a multi-agent coding harness that scales to 1,024 concurrent agents with no central orchestrator directing them. On the Pandoc rebuild benchmark, test-pass rates climbed from 33.89% with a single agent to 55.06% with 1,024, all under the same six-hour budget. No boss. Better results. That is the entire argument against the way most of us build multi-agent systems today.
The Orchestrator Is a Bottleneck by Design #
Every mainstream multi-agent framework — LangGraph, CrewAI, Microsoft’s own Agent Framework — routes coordination through a central orchestrator. That orchestrator decides who does what, handles dependencies, and merges results. It works at small scale. At 50 or more agents it starts spending more tokens on coordination than on the actual task, must understand every domain it oversees, and becomes a single point of failure for the entire system.
Agensh’s answer: remove the orchestrator entirely. Each agent reads the same shared state, claims its own work, and merges its own code. The “manager” is a shared task board, not another agent.
How Agensh Coordinates Without a Boss #
The system runs on three coordination planes:
- Shared workspace (Gitea) — A Git repository where agents develop contributions in parallel, resolve merge conflicts, and publish completed work. Concurrent writes, full version history.
- Message interface (Mattermost) — A shared channel for team-wide announcements and direct messages for urgent peer coordination. Lower-priority channel messages land at loop start; high-priority DMs arrive at each tool call.
- Shared context (board service) — Stores reusable findings across the whole organization, so agents build on each other’s work without re-discovering what a peer already learned.
The cooperation loop itself lives in the prompt, not in runtime code. Every worker gets the same prompt — only its ID differs. Each agent independently gathers context, claims a sub-task, takes action, shares findings, verifies results, and merges progress before starting the next iteration. No routing agent. No central state machine. The system is the shared state.
What Emerges When No One Is in Charge #
The results across five ProgramBench tasks: mean test-pass rate goes from 19.31% at one agent to 28.78% at 128 — a 49% relative gain. But the more interesting finding is what happened as the teams grew larger: agents spontaneously developed specialized roles. Some drifted toward integration work, others toward exploration. Peer-verification workflows emerged without being programmed. Agents began filing direct messages when they detected work overlap.
None of that was coded. It came from the structure of shared state plus same-prompt design. The implication is significant: Agensh challenges the model of explicitly assigning agent roles. When you give agents the same instructions and a shared environment, role differentiation appears on its own.
You Can Run It Today #
Requirements: Linux with Docker Engine, Compose v2, and Python 3.11+. The default configuration spins up 8 workers on a single node with a 6-hour task budget. The 1,024-agent run used 16 nodes at 64 agents each, with staggered activation — one new agent every 30 seconds for the first hour — to reduce task contention during cold start.
The release includes the full harness, experiment launch scripts, and a reproduction guide. You can run the single-node baseline on a workstation today; scaling to multiple nodes requires coordinating Gitea, Mattermost, and the board service across machines. The defaults use GPT-5.6-sol at high effort, but the model is configurable.
One Microsoft, Two Agent Philosophies #
Here is the awkward subtext: Microsoft’s own production multi-agent story — the Agent Framework, which went GA in April 2026 as AutoGen’s successor — is orchestrator-based. It routes decisions through a central coordinator and is Microsoft’s endorsed path for enterprise teams building on Azure.
Agensh is a research project going in the opposite direction. Microsoft Research is running 1,024 agents with no central controller and getting better benchmark numbers than the orchestrator model produces at scale. Both live at Microsoft. Neither team seems bothered by the contradiction — which suggests this is a genuine open research question rather than a settled architecture decision.
If the Agensh model holds up at production scale — handling real-world task complexity beyond ProgramBench — it would represent a meaningful architectural shift for anyone building multi-agent systems. The benchmark evidence is compelling enough to watch closely.