Impressive NVIDIA paper on cutting agent costs.

For long-horizon agent tasks, you want the capable model for the early steps and a cheaper model for the rest. Handing off in the other direction saves little or nothing.

AgentBaton learns when to make that one switch from patterns in task progress. Each check has a median latency of 338 ms on one H100.

Across 11 model pairs and two benchmarks, it cuts API cost by 65.9% on average while largely keeping accuracy. Some pairs gain up to 8.8 points over running the capable model alone.

The handoff happens once per task, so it is easy to add to an existing agent loop.

If your agent bills are dominated by a frontier model, you might want to test one capable-to-cheap handoff before building a full router.

https://arxiv.org/abs/2610.35959

Chat with Paper: https://academy.dair.ai/papers/agentbaton-efficient-model-collaboration-in-agentic-tasks-2610.35959