Strong AI agents hand the move to a human in 80% of cases when the human's path is shorter

If an agent asks a human for help less and less often, that does not necessarily mean the collaboration is falling apart. The work may simply have been divided correctly. A new study from Hugging Face Daily Papers tests this hypothesis in a maze with two avatars. Four language models, GPT 5.6 Sol, Command A+, Qwen 3.8 27B and Command R, played alongside humans. The models that reliably solve the task on their own handed the human roughly 80% of first decisions when the human's route was shorter. When the human had no advantage, they almost always acted themselves.

A maze where you can hand off the move

The authors start from a simple observation: the share of human participation cannot tell bad collaboration from a good division of labor. An agent that doesn't call on the human may be overconfident, or it may simply be doing better. To separate these cases, you need to control who is stronger at the task.

For this, the researchers built a shared maze. It has two avatars, one controlled by the agent and the other by a human. The goal is common: reach a single point in the minimum number of team moves. On each turn, the agent chooses whether to act itself or pass the move to its partner.

The human's advantage is set only by the starting layout, that is, by the length of the human's shortest path. Everything else in the task stays the same. This produces three conditions: the human has a clear advantage, neither side has an advantage, or the agent has the advantage.

First solo, then as a pair

The experiment ran in two stages. First, each model was tested alone: one avatar, with the map fully observable. This measured the model's raw ability to solve the maze without a partner. Then the models played in pairs with humans under all three conditions.

The main result: the better a model performs alone, the more selectively it involves the human, depending on the route. The key figures:

• Models that reliably solve the maze themselves act on their own almost without exception when the human has no advantage.
• When the human's route is shorter, these same models hand over about 80% of initial decisions.
• Command R, which is noticeably weaker in solo mode, did not show this selectivity.
• Human intervention changes with the task condition rather than staying a fixed habit of the model.

The finding on Command R deserves separate attention. Its behavior does not look like a deliberate allocation of roles. Judging by the results, the ability to tell whose path is shorter is linked to the ability to solve the task. A model that builds routes poorly for itself also judges its partner's route poorly.

Why the “how many times it called the human” metric is misleading

For the agent industry, this is bad news about familiar metrics. Many teams track the rate of escalations to a human and try to push it down. Here, a lower rate can signal either maturity or blindness. An agent that fails to hand over the move when the human is faster loses efficiency. An agent that hands over the move indiscriminately offloads work it could have done itself.

So what matters in this study is not how often the agent asks for help, but whether its requests match the conditions. A good agent behaves differently in different configurations: it takes the task on itself when it is stronger and hands it off when the partner is closer to the goal.

What to change in agent evaluation

For teams building human-in-the-loop agents, the practical takeaway comes down to a few steps.

• Don't measure the overall escalation rate. Measure how escalations depend on who is objectively stronger at a given subtask.
• Build tests with a controlled advantage for the human, for the agent, and with zero advantage, as in this maze.
• Measure the model's solo ability first, and only then interpret its behavior in a pair. Without this baseline, you can't tell whether the agent delegates deliberately or blindly.
• Don't treat a weaker model as more “cooperative” just because it asks for help more often. Command R shows no selectivity, and its request rate says nothing on its own.

The maze remains a lab task, and the results should be carried over to real workflows with caution. But the method is useful: it separates the question “can the agent work with a human” from the question “can it solve the task itself.”