Start capable.
Finish efficient.
Capable-to-efficient handoffs generally offer more favorable accuracy–cost trade-offs than the reverse direction.
AgentBatonEfficient Model Collaboration
in Agentic Tasks
A capable model starts. An efficient model finishes.
Learn when to pass the baton.
♣ The Hong Kong University of Science and Technology♠ University of Illinois Urbana-Champaign
See it in action
Same task. A different handoff.
The efficient model fails alone. With the capable model’s early progress, it completes the task at a fraction of the accounted API cost.
Qwen aloneFailed$0.00 accounted API cost
Opus alonePassed$7.02 accounted API cost
AgentBatonPassed$1.13 accounted API cost
Recorded Testem trajectory replay, not live inference. Handoff at turn 13. Local-model compute and detector costs are excluded; this example illustrates cost savings, not faster execution.
The question
Can models collaborate within one task?
Solving long-horizon agentic tasks requires language agents to interact with their environments over multiple steps, and reliable performance often comes from the expensive capable models available through cloud providers. Meanwhile, alternative efficient models with inferior performance but a much lower cost are emerging, including lower-tier cloud models and smaller open-weights models deployable on local hardware. This raises a natural question that remains less explored than traditional query-level routing: can capable and efficient models collaborate within a single long-horizon agentic task to substantially reduce cost while preserving task accuracy? In this paper, we systematically investigate key design choices for such collaboration through model handoffs, where one model takes over from another, focusing on handoff direction and context transfer. We find that efficient-to-capable handoffs generally underperform capable-only execution with limited or even no cost reduction, whereas capable-to-efficient handoffs at suitable stages of task progress largely preserve or even improve accuracy while substantially reducing cost. These findings suggest that a one-time, simple, well-timed handoff strategy is good enough for cost-effective agent collaboration, motivating AgentBaton, a lightweight detector that learns when to hand off from recurring patterns in task progress. Across 11 model pairs and two representative benchmarks, AgentBaton reduces API costs by 65.9% on average while largely preserving task accuracy, with accuracy gains of up to 8.8 percentage points over capable-only execution. Experiments across various reasoning-effort settings further show that AgentBaton extends the empirical accuracy–cost Pareto frontier.
The handoff structure
One task, two models, one handoff
A source model begins a task and transfers control once to a receiving model. The receiving model inherits the workspace and continues with the prior interaction history, supplied in context or through log files.
Contribution 01
Understanding when handoffs work
Using this setup, we systematically compare handoff direction, timing, and context transfer. The findings favor a simple, well-timed transition from a capable model to an efficient one.
Start capable.
Progress matters
more than a fixed turn.
A useful implementation and informative feedback can narrow the remaining work enough for an efficient model to take over.
Transfer what
the next model needs.
The best choice between in-context history and log files depends on the source model and task stage.
Contribution 02
Learning when to pass the baton
The study reveals recurring signs of handoff readiness: a useful implementation, informative execution feedback, and more focused remaining work. These findings motivate AgentBaton, a lightweight detector trained to recognize when an efficient model can take over.
Online handoff detection
A small detector.
A single decision.
- Observe task progress
Append the latest action and tool response to the detector’s growing history.
- Continue or hand off
Not ready: keep using the capable model. Ready: transfer control to the efficient model.
- Complete the task
The efficient model continues with the workspace and context already established.
2B parameters4,000 training trajectories338 ms median latency on H100
Results
Extending the accuracy–cost frontier
Across 11 model pairs on DeepSWE v1.1 and Terminal-Bench 2.1, learned handoffs reduce mean API cost by 65.9%, with accuracy outcomes varying by configuration.
Kimi K3 → Qwen 3.8 27B
72.9% lower cost
72.6% accuracy, compared with 73.5% for capable-only execution on DeepSWE.
Claude Opus 5 → GPT-5.6 Luna
67.3% → 76.1%
Higher task accuracy alongside 70.5% lower API cost on DeepSWE.
Citation
BibTeX
@article{li2026agentbaton,
title = {AgentBaton: Efficient Model Collaboration in Agentic Tasks},
author = {Li, Junlong and Jiang, Jindong and Fu, Yonggan and Yao, Jiarui
and Ye, Hanrong and Byeon, Wonmin and Belcak, Peter
and Zhang, Hao and Molchanov, Pavlo},
journal = {arXiv preprint arXiv:2610.35959},
year = {2026},
url = {https://arxiv.org/abs/2610.35959}
}