It’s Rational for AI Agents to Procrastinate

We propose labour games, a class of unmediated mixed-motive cooperation games where rational agents must collectively complete a task by taking a negative-reward labour action $U$ times over $T$ timesteps. Each agent aims to maximize their utility by minimizing their contribution, achieving an optimal balance of labour and completion chance. We first study a single agent in a system of opaque agents, and find a tendency for delaying labour actions, dubbed critical completion, similar to procrastination. Using this result, we investigate the multi-agent system, defining the rules governing the general behaviour of agents in a labour game. We find that agents have a preference for contiguous action leading to critical completion and characterize this behaviour as the state of commitment which we show is common knowledge. We show that commitment prevents less capable rational agents from contributing in labour games and illustrate the link between commitment and emergent communication. With these principles as a guide, we propose a mechanistic mitigation to the undesirable phenomena and analyze its effects on the multi-agent system. We experimentally illustrate the discoveries of our theoretical analysis using a multi-agent reinforcement learning encoding of the game. Finally, we provide a discussion on the implications of our results, the inefficacy of current reward structures for practical multi-agent systems, and the open question of how to build appropriate incentive structures for productive agents.