AI Papers Reader

Personalized digests of latest AI research

View on GitHub

Who Takes the Lead? Capability and Task-Based Collaboration in AI Agents

As language-model agents become more capable, they can finish more work without asking for help. A decline in how often an agent involves a human partner might look like poor collaboration, but researchers led by Weien Li and Ye Yuan of McGill University argue it could instead be a sensible division of labor. Their study suggests that involvement rates alone cannot tell the two situations apart.

To separate them, the team built a “shared-control maze,” a grid puzzle in which an agent-controlled avatar and a human-controlled avatar share one goal: reach a common target in as few team turns as possible. On each turn, the model either moves its own avatar, directs the human with a recommended direction, or hands the move to the human without any recommendation. The model sees the full maze layout, while the human sees only a local view. The researchers set the starting positions so that the agent’s route was shorter, the routes were tied, or the human’s route was shorter. The maze and objective stayed the same, and the model was never told which route was shorter.

A concrete case shows the mechanism. When the human’s avatar started with a five-move route and the agent’s with a seven-move route, GPT 5.6 Sol directed the human on 82.5% of initial decisions, but it never moved its own avatar only when the agent’s route was shorter or tied.

Before testing collaboration, the researchers measured each model working alone. Qwen 3.8 27B, Command A+, and GPT 5.6 Sol solved the mazes reliably, with Qwen and GPT 5.6 Sol at 98% and 100% of episodes completed within ten turns, respectively. Command R completed 22% of episodes.

In co-play, the three stronger models showed a clear pattern. When the agent’s route was shorter or tied, they took the first move themselves almost every time. When the human’s route was shorter, Qwen 3.8 27B and GPT 5.6 Sol involved the human on 82.5% of initial decisions, and Command A+ did so on 77.5%. Command R showed no such pattern. It took the first action itself 70% of the time when the agent’s route was shorter, 45% when routes were tied, and 57.5% when the human’s route was shorter. Its overall action mix also barely changed with route advantage. Human involvement took the form of guidance almost entirely. No model deferred without a recommendation on an initial decision.

The study has clear limits. It tested four models, four human participants (all authors), and one maze environment with 5×5 grids. The authors describe their findings as an association between solo ability and selective involvement, not proof that greater capability causes it. Command R’s completion rate also rose when the human started closer, but because the experiment assigned route advantage rather than involvement, the authors do not claim involvement caused the improvement.

The work offers a practical idea for evaluating AI teammates: a low rate of human involvement is not automatically a flaw, and a high rate is not automatically a virtue. Whether the approach extends to richer tasks, such as coding or everyday work, remains an open question.