AgentBaton Lets Capable and Efficient Models Take Turns in Agentic Tasks

Long-horizon agents do more than answer a single prompt. They plan, interact with tools and environments, observe the results, and make decisions repeatedly. Running an expensive, highly capable model through every step can improve reliability, but it also makes the cost of an agentic workflow grow quickly. Smaller open-weight models and lower-tier cloud models are cheaper, yet may struggle with the hardest parts of a task.

AgentBaton examines whether the two classes of models can collaborate within one continuous task. Its focus is not conventional query-level routing, where every request is assigned independently. Instead, one model runs for part of the trajectory and then hands the task to another model, which continues from transferred context.

Key findings

• Handoff direction matters. Moving from an efficient model to a capable model generally performs worse than using the capable model throughout, while providing limited or no cost benefit. The reverse direction is more promising: once a capable model has completed important reasoning or established task progress, an efficient model can often continue at lower cost.
• A single handoff may be enough. The study does not require constant switching between models. Its results suggest that a simple handoff at an appropriate stage can capture much of the benefit without adding a complex orchestration layer.
• AgentBaton detects the transition point. The proposed lightweight detector learns recurring patterns in task progress and decides when control should move from the capable model to the efficient one.
• The trade-off improves in the reported tests. Across 11 model pairs and two representative benchmarks, AgentBaton cuts API costs by 65.9% on average while largely preserving task accuracy. In some settings, it improves accuracy by as much as 8.8 percentage points compared with capable-only execution. Experiments with different reasoning-effort settings also place the method farther along the empirical accuracy-cost Pareto frontier.

Why it matters

The practical message is that a strong model may not need to own an agentic workflow from beginning to end. It can handle planning, difficult decisions, or high-risk recovery steps first. After the trajectory becomes more structured and easier to continue, a cheaper cloud model or a locally deployed model can take over. This could reduce recurring inference expenditure and lower dependence on premium hosted models.

The strategy is not automatic for every workload. A handoff that happens too early may leave the efficient model with unresolved reasoning, while one that happens too late gives up most of the possible savings. Context transfer must also preserve the state needed for reliable continuation. The available material does not establish that the same policy will work across all agent environments, so the results should be read as design guidance rather than a universal routing rule.

AgentBaton’s broader contribution is a change in framing: model selection is not only about which model answers a request, but also about when a model should give up control. For long, cost-sensitive workflows with distinguishable phases, this lightweight form of model collaboration is a useful direction for agent orchestration.

Source: Hugging Face Daily Papers