Abstract
When agents collaborate with unfamiliar teammates, they must decide not only how to act, but also when to rely on others. This is challenging because a teammate may be strong on some subtasks but less effective on others, making delegation both useful and risky. Delegating control can improve performance when the teammate has the right skill, but it can also waste time or reduce performance when the teammate is less competent. Delegation also provides useful evidence about the teammate’s competence through their success or failure on the current subtask. We study this tradeoff in ad hoc teamwork, where an agent must learn when to act on its own and when to delegate to an unseen teammate whose strengths vary across subtasks. We propose CoDeT, a competence-belief delegation method that uses factored Bayesian beliefs to estimate teammate competence across subtasks and decide when to delegate. We evaluate CoDeT using a strict complementarity metric, which measures whether teamwork creates value beyond what either agent could achieve alone. Across sequential decision making tasks including Level-Based Foraging and Overcooked, CoDeT achieves positive teamwork value, avoids harmful delegation, and outperforms baselines that do not learn when to delegate. In settings where teammate contributions are consistently useful, CoDeT also recovers delegation strategies that are competitive with strong static engagement policies. These results suggest that factored competence beliefs can support more calibrated delegation, with implications for human-AI and multiagent teamwork.
Funding Information
  • National Science Foundation (2451134)