Persuasion Propagation in Long-Running AI Agents

A new study from arXiv (2602.00851v3) investigates how user persuasion affects AI agents engaged in long-horizon tasks like coding and web research. The researchers introduce the concept of "persuasion propagation," where belief-level intervention influences downstream task behavior. They developed a behavior-centered evaluation framework distinguishing persuasion applied during task execution from that applied prior. Results show that on-the-fly persuasion has weak and inconsistent effects, but when belief states are explicitly specified at task time, belief-prefilled agents show significant behavioral changes—conducting on average 26% more actions. The study highlights challenges in studying long-running agent behavior due to noise and cost.