Habr AI: Self-Organizing LLM Agents Outperformed Hierarchical Systems by 14%

An experiment with LLM agents showed that conventional organizational logic transfers poorly to AI systems. If you don't impose roles on agents from above, but let them choose their specialization and level of participation themselves, solution quality improves.

How the Hypothesis Was Tested

The researchers spent six months testing different coordination schemes on 25,000 tasks, using 8 models and teams of up to 256 agents. The main question was simple: does the same approach that works for people apply to AI—one with a coordinator, fixed roles, and predetermined structure? To verify this, they compared several operating modes—from rigidly designed teams to systems where agents themselves decide who and when to engage with a task.

In the self-organizing scheme, an agent doesn't receive a label like analyst, editor, or reviewer before work begins. Instead, it looks at the specific task, assesses where it can be useful, and chooses specialization based on the situation. Moreover, an agent can opt out entirely if its contribution won't improve the outcome. This is an important shift: instead of discipline and obedience, the system relies on local decisions from each participant and assembles them into an overall strategy.

Why Hierarchy Lost

The key finding of the study sounds harsh: assigning roles in advance is an antipattern. The system where agents independently distributed functions outperformed the coordinator variant by 14%. The reason isn't just flexibility. When a role is predetermined, an agent starts fitting its behavior to the template, even if the task requires a different type of thinking. As a result, some participants work not where they provide maximum value, but where an architect once placed them.

Assigning roles is an antipattern.

The most telling result isn't about percentages, but about behavioral diversity. Just 8 agents created 5,006 unique roles during the experiment—far more than a person typically builds into such a system's design. This isn't chaos, but dynamic microspecialization: the same agent can search for facts in one case, clarify requirements in another, and remain silent in a third. The very right to not participate, when the value of contribution is low, separately boosted the quality of final answers.

Practice for Teams

For developers of multi-agent systems, this implies a fairly practical set of rules. If the task changes from case to case, a rigid organizational structure begins to hinder rather than help. Instead of complex hierarchies, it's more useful to design selection mechanisms: who takes the task, how an agent signals its competence, and when it should exit the process. Otherwise, a team of agents quickly becomes a digital copy of a corporate department with all its unnecessary approvals.

• Don't fix roles where tasks differ significantly from each other
• Give agents the ability to choose their own specialization for a specific request
• Allow them to opt out if confidence is low or the contribution would be redundant
• Evaluate not only system obedience, but also the quality of self-organization
• Design scaling from small groups to large ones without changing the coordination principle

The practical value of the work is that it doesn't reduce to a beautiful theory about emergent behavior. It provides direct recommendations for those building AI pipelines, assistants, and agent platforms: less manual management, more rules for local choice. This approach is especially important where tasks are heterogeneous, context changes quickly, and the cost of an extra step is high. Under these conditions, self-organization turns out to be not a research exotic, but a way to get better results with the same set of models.

What This Means

For the market, this signals that the next stage of agent system development is linked not to increasing hierarchy complexity, but to designing an environment where agents can negotiate without a rigid boss. If the findings from the experiment are confirmed in applied products, many teams will reconsider the architecture of AI assistants: from a set of pre-assigned roles to more flexible, adaptive, and economical coordination.