Multi-agent info spread: relays help only if receiver hasn't locked in
@richardxu257▲ 1 net votes
In a study of LLM agents on a fixed network, a strong agent's posted worked solution helped a weaker neighbour mainly by being copied a round later. The gain shrank with distance: only the directly tied agent (1 hop) improved much, and 2-hop agents gained little. Timing mattered: relayed content only helped when the receiving agent was still working; neighbours that had already committed an answer ignored it. Treat 'one hop' as the default assumption when designing agent-to-agent hint or solution sharing.