Agensh Scales Self-Organizing Collaboration to 1,024 Agents

Introduction

The appeal of multi-agent systems is not simply that more models can work at once. Their real promise is to divide a difficult job into parallel pieces and reduce the time required to complete it. Yet many existing designs depend on a central orchestrator to assign work, monitor progress, and resolve coordination issues. As the number of workers grows, that coordinator can become the next bottleneck. Agensh takes a different route: it removes the central controller and allows the organization to coordinate itself.

A lightweight infrastructure for local decisions

Agensh does not rely on a fixed hierarchy in which every agent receives a predefined role. Instead, concurrent workers repeatedly follow a cooperation loop. They gather the latest context, identify or claim a subtask, decide what action to take, share discoveries, verify available results, and merge valid progress into the broader effort. Because the loop is asynchronous, workers do not have to wait for a central service to issue instructions one by one.

The system rests on three components:

• Shared workspace: tracks proposed, ongoing, and completed work so agents can discover tasks that need attention.
• Message interface: allows workers to exchange findings, report progress, and communicate decisions.
• Shared context: stores reusable discoveries and work intentions, helping later agents understand the current state and avoid unnecessary repetition.

The key design choice is to make project state visible enough for coordination to emerge from local actions, rather than adding another increasingly complex management layer.

Evidence from programming tasks

The paper evaluates Agensh with GPT-5.6-sol (high) on the five hardest ProgramBench tasks. Increasing the organization from one to 128 agents raises the mean final test-pass rate from 19.31% to 28.78%, an approximately 49% relative improvement. The researchers also report that larger organizations often reach comparable pass rates earlier, suggesting that parallelism can improve both execution speed and final outcomes.

The pandoc result provides a more expansive scaling example. Moving from one agent to 1,024 agents raises the final test-pass rate from 33.89% to 55.06%. Analysis of worker trajectories further suggests that different forms of self-organized cooperation gradually appear and become more standardized as the organization expands.

Why it matters—and what remains open

Agensh points to an alternative scaling principle for agentic systems. More capability does not necessarily have to come from a stronger central planner; shared state, communication, and local autonomy may also produce useful organizational intelligence. For long-running programming tasks and other problems that can be decomposed, this could reduce pressure on a coordinator and expose more opportunities for parallel work.

The evidence, however, comes from a specific programming benchmark and model configuration. It should not automatically be generalized to every workload. A larger population also creates more communication, state-management, and verification overhead. Controlling duplicated effort, conflicting information, and operating cost will be essential if self-organizing agent systems are to move beyond experimental settings.

Source: Hugging Face Daily Papers