@Max Planck Institute for Intelligent Systems

Current AI agents are fundamentally limited because they operate on a single, sequential token stream, forcing them to perform reading, thinking, and acting one step at a time. Researchers from the Ma…