Microsoft Research has studied the swarm behavior of AI agents

Researchers from Microsoft Research conducted an experiment on swarm interaction of language models without centralized control. As part of the Agensh preprint, up to 1024 instances of a single model spent six hours collaboratively reproducing the behavior of finished programs in a shared repository.

During the experiment, the AI agents worked under conditions of complete task equality, without an assigned team lead or manager. The autonomous system independently distributed roles, identifying reviewers and integrators among the participants, and established internal protocols for coordinating actions.

According to the test results, the collective of multiple agents demonstrated higher task execution efficiency compared to single runs of similar models. At the same time, the researchers recorded specific swarm behaviors during large-scale parallel coding.

The data obtained allows for evaluating the scalability of autonomous systems in software development without human involvement in subtask distribution.