Coordinating Human-AI Teams for Enhanced Performance

The increasing capabilities of AI agents necessitate a deeper understanding of how humans and AI can collaborate effectively, especially in shared workspace environments where tasks require both AI efficiency and human judgment. This study explores the conditions under which adding AI or human collaborators improves team performance versus creating coordination overhead. Using the Collaborative Gym environment, the research found that simply adding more collaborators, even relevant ones, can sometimes degrade performance if the team lacks structured coordination. This "process loss" occurs when the benefits of additional expertise are outweighed by the challenges of integrating contributions. To address this, the study evaluated scaffolding mechanisms, specifically combining shared group memory with human-in-the-loop (HITL) gates. These gates require human approval for selected AI actions. This structured approach led to higher mean performance, particularly in three-person teams, by providing clearer responsibility signals and better routing of expertise, emphasizing that effective coordination is as vital as individual capabilities.