Agensh: Scaling Organizational Intelligence to 1,024 Agents

Agensh. Scaling Organizational Intelligence to 1,024 Agents. By Zhihao Zhan and six others, from Microsoft Research. Published on the twenty-second of September. It links to a code repository that is no longer there.

A multi-agent system can cut latency on complex tasks by doing work concurrently. Several pioneering harness frameworks support such systems. But their scalability is often limited by a central orchestrator, and by how much that orchestrator can handle in allocating tasks and coordinating workers. As the number of workers grows, the orchestrator becomes the bottleneck. Every allocation and every coordination decision passes through it, so the capacity of that one component caps the whole system.

Agensh removes the orchestrator. It is a scalable, self-organized multi-agent harness in which concurrent workers each run a cooperation loop. They continuously gather context, claim and self-assign sub-tasks, act and share what they find, verify results, and merge progress, all asynchronously. Nobody hands out the work. The workers take it.

The loop is supported by what the authors call agentic organization infrastructure, which has three components. A shared workspace holds work that is proposed, ongoing and completed. A message interface lets workers talk to each other. And a shared context keeps reusable findings and records of what each worker intends to do. Between them, these three pieces do what an orchestrator would otherwise do. The workspace shows what needs doing and who is on it. The messages let workers coordinate directly. The shared context stops them from repeating each other's discoveries or colliding on the same work.

To test scalability, Agensh was run on the five hardest ProgramBench tasks with GPT-5.6-sol at high setting. Scaling from one agent to one hundred and twenty-eight raises the mean final test-pass rate from nineteen point three one percent to twenty-eight point seven eight percent, a relative improvement of about forty-nine percent. Larger organizations also reach comparable pass rates earlier. So adding agents helps in two ways: the final result is better, and good results arrive sooner, which is the latency benefit a multi-agent system was meant to deliver in the first place.

On pandoc, the scale goes further. Moving from one agent to one thousand and twenty-four raises the final test-pass rate from thirty-three point eight nine percent to fifty-five point zero six percent. Worker trajectories show different forms of self-organized cooperation emerging gradually and becoming standardized as the organization grows. The structure is not designed in advance. It appears as the number of workers rises.

The authors read these results as establishing the number of agents as a new scaling dimension for multi-agent organizations. With no central orchestrator in the loop, adding workers kept helping, up to one thousand and twenty-four of them on pandoc. They offer it as a practical option for complex tasks under hard limits on latency or time.