Escher-Loop: Mutual Evolution by Closed-Loop Self-Referential Optimization

ArXi:2604.23472v1 Announce Type: new While recent autonomous agents nstrate impressive capabilities, they predominantly rely on manually scripted workflows and handcrafted heuristics, inherently limiting their potential for open-ended improvement. To address this, we propose Escher-Loop, a fully closed-loop framework that operationalizes the mutual evolution of two distinct populations: Task Agents that solve concrete problems, and Optimizer Agents that recursively refine both the task agents and themselves.