FudanNLP work that automatically evolves coding-agent harnesses from observability traces — mining execution logs to improve the scaffold, with transfer across model families. Part of the 2026 wave treating the harness (not just the model) as the object of optimization.

Paper

agentsinfrastructureresearch