TL;DR: Let a frontier model do the heavy lifting, then hand the baton to a cheaper model once only finishing work remains. API cost drops 65.9% on average with accuracy largely preserved.
AgentBaton: Efficient Model Collaboration in Agentic Tasks
Key points
・🔀 Instead of per-query routing, control moves once mid-task from a capable to an efficient model: a "handoff"
・🧭 The direction matters: cheap→capable handoffs help little, capable→cheap works
・⏱️ Good timing: an artifact exists, feedback has localized the problem, remaining work is local. Just prompting for it is unreliable; Kimi K3 never handed off (0/113 tasks)
・🤖 A lightweight detector initialized from Qwen 3.5 2B judges each turn, with 338ms median latency
・💰 On DeepSWE, Kimi K3→Qwen hits 72.6% (vs. 73.5% alone) at 72.9% lower cost
・📈 Claude Opus 5→GPT-5.6 Luna rises from 67.3% to 76.1% at 70.5% lower cost
・🆚 Versus the per-turn router MTRouter: 46.7–86.6% cheaper with equal or better accuracy
In agent operations, when to switch models may matter more than which model to use.
#
AIAgents# #
LLM#