Recent work has shown that LLM agents can evolve their own harness, up to and including the team itself. But a new arXiv preprint suggests that this self-evolution does not happen in a vacuum. Borrowing the notion of scaffolding from human social groups, the authors argue that the surrounding social structure plays a critical role in shaping how agents develop and perform.

The key claim is that identical agent abilities can produce different outcomes depending on the social conditions in which the agents operate. In human groups, the same skills lead to different results under different social arrangements, and the paper suggests the same holds for multi-agent systems. This points to a need to design the social context alongside the agents themselves.

The abstract is brief and does not yet detail the specific scaffolding mechanisms or experimental results. But the framing suggests a shift from optimizing individual agents to engineering the social environment that supports their evolution.