Workspace Optimization: Training Agents Without Weight Updates

A new paper on arXiv proposes workspace optimization as a method for training language model agents without modifying their weights. The approach focuses on evolving the agent's external workspace—the structured substrate it reads, writes, and tests—through interaction. The authors introduce DreamTeam, a multi-agent system for ARC-AGI-3 that uses roles to build world models, plan, hypothesize, and route failures. The method mirrors weight-space training with artifacts as parameters, evidence as data, counterexamples as losses, and textual feedback as gradients. The paper targets hard multi-turn environments where frontier models have strong priors but cannot solve tasks in a single shot.