Tag
This paper systematically examines how manipulating personality traits in multi-agent LLM teams affects performance across coding, research collaboration, and bargaining tasks, finding that effects depend critically on task structure.
The article points out that behind Claude Code's ability to automatically generate workflows, it reflects that the control plane of AI agent products is shifting from relying on long contexts to remember goals and decompose steps, towards externalizing into an executable harness, including task structure, permission boundaries, verification mechanisms, and stop conditions.
This paper establishes nonparametric identifiability guarantees for extracting task-relevant representations from generalist models, proving that task structure is identifiable across time steps and latent representations are identifiable within each step under sparsity regularization.