A difficult task rarely stays equally difficult from beginning to end. Finding the right approach may require one kind of model; checking and finishing an established implementation may require another.
Routing inside a trajectory
AgentBaton studies handoffs within an agentic task. Its empirical finding is directional: starting with a capable model and later transferring to an efficient model can offer a better trade-off than the reverse.
A lightweight detector learns to recognize handoff readiness from the task and the observed trajectory. The method section describes supervision based on task progress, including an existing substantive artifact, concrete execution feedback, and remaining localized work. The detector is initialized from Qwen 3.5 2B.
Think in phases, not fixed turn counts
Consider an illustrative debugging task. During the early phase, an agent must identify the failing component and select an approach. Later, after producing a patch and observing a precise failing assertion, the remaining work might be a small local correction.
A rule such as “switch after ten turns” ignores this difference. Ten turns might represent successful diagnosis in one task and repeated confusion in another. A progress-sensitive decision instead asks what state has been established and whether the next model can use it.
The receiving agent also needs the relevant context. A correct handoff time is not helpful if the evidence required to finish disappears during transfer.
The reported trade-off
Across 11 model pairs and two benchmarks, the paper reports 65.9% average API-cost reduction while largely preserving accuracy, with gains of up to 8.8 percentage points in particular settings. Those aggregate figures do not imply that every pair preserves quality: the detailed results include accuracy regressions as well as improvements.
How to evaluate a handoff policy
A useful study would report unsuccessful handoffs, detector overhead, context-transfer cost, and final success together. Otherwise, low token cost can obscure extra recovery work. It is also useful to test unfamiliar tasks, since recognizable progress in one domain may not mark readiness in another.
This complements Nemotron 3.5: one concerns what a model learns, the other how models share an execution budget.