AgentBaton: Efficient Model Collaboration in Agentic Tasks

AgentBaton: Efficient Model Collaboration in Agentic Tasks

Long-horizon AI agents can be expensive to run. A capable cloud model may need dozens of tool-using turns to inspect files, write code, test changes, and respond to failures. Smaller or lower-cost models can reduce the bill, but they often struggle when they must solve the entire task from scratch.

AgentBaton, from Junlong Li and colleagues, studies whether these models can collaborate during one task rather than being selected only at the beginning. Its central idea is a carefully timed handoff: let a capable model establish meaningful progress, then transfer control to a more efficient model for the remaining work.

The problem with routing by query

Traditional model routing often chooses a model from the initial request. That assumption is less reliable for interactive agents. Actions change the workspace, tool feedback reveals new information, and the difficulty of the remaining work can differ substantially from the difficulty suggested by the original prompt.

The authors therefore examine two handoff directions. In an efficient-to-capable handoff, a cheaper model starts and a stronger model takes over. In a capable-to-efficient handoff, the stronger model establishes the solution and a cheaper model finishes it. The study also compares two ways to transfer history: placing the previous interaction directly in the receiving model’s context, or saving it in a log file that the receiving agent can inspect when needed.

What AgentBaton does

The proposed system uses a lightweight detector that evaluates the capable model’s trajectory after each turn. It looks for signs that the task has crossed a useful transition point. The receiving model inherits the current workspace and continues until completion; the system does not repeatedly switch models back and forth.

The training labels reflect three broad conditions identified by the authors’ trajectory analysis:

  • A substantive implementation, prototype, or task-required artifact exists.
  • Execution, testing, or inspection has produced task-relevant feedback.
  • The remaining work is mostly localized repair, diagnosis, validation, or finalization rather than choosing and implementing the core approach.

This distinction matters. A failed test can still make a handoff appropriate if it narrows the problem. Conversely, a plan or untested code sketch is not enough. In the authors’ examples, the capable model might fix a quoted-path bug, validate a coordinate transformation, or localize an encoding fault. Once that groundwork is present, the efficient model can often complete the task.

Evidence from two agent benchmarks

Li and colleagues evaluate the approach on Terminal-Bench 2.1 and DeepSWE v1.1, using multiple capable models and efficient receiving models. Their broader evaluation covers 11 model pairs. The headline result is a 65.9% average reduction in API cost while largely preserving task accuracy. The paper also reports accuracy improvements of up to 8.8 percentage points over capable-only execution in its evaluated configurations.

The fixed-checkpoint experiments reveal a clear asymmetry. Starting with an efficient model and escalating later generally does not deliver substantial savings without an accuracy penalty. In some cases, the stronger model must spend effort recovering from an earlier trajectory. By contrast, capable-to-efficient handoffs can work well once progress is sufficiently mature.

For example, on DeepSWE, a Kimi K3 trajectory handed to Qwen 3.8 27B at a suitable stage reached 72.6% accuracy versus 73.5% for Kimi alone, while reducing cost by 72.9%. In another reported comparison, GLM-5.3 handed to Qwen at the detector’s chosen point reduced cost by 84.5%, although accuracy was lower than the capable-only baseline. The results are therefore encouraging, but not uniformly lossless across model pairs and benchmarks.

The authors report that 12 of 22 model-pair and benchmark settings either improved accuracy or lost less than 5% relative to capable-only execution. AgentBaton also extended the empirical accuracy-cost Pareto frontier in reasoning-effort sweeps. In one example, a Kimi configuration handed to GPT-5.6 Luna achieved 79.6% accuracy at $1.60 per task, exceeding the best single-model accuracy in that sweep at a lower reported cost.

Context transfer is not one-size-fits-all

The experiments do not identify a universally best history format. Log files generally helped when an efficient model handed off to a capable one, often reducing context-related cost. For capable-to-efficient handoffs, in-context transfer worked better for some source models, while logfile transfer worked better for others. The authors suggest that differences in available reasoning traces may partly explain this variation.

This is a practical warning against treating “handoff” as a complete design. The receiving model needs access to the workspace and enough information to understand what has already been tried. Giving it everything may be expensive or bias it toward unproductive assumptions; giving it too little may force redundant investigation.

Limitations and practical implications

The paper’s strongest claims are empirical, not universal. The experiments focus on two benchmarks, a simple agent scaffold, and one-time handoffs. Results vary by model, transfer method, handoff timing, and benchmark. The cost accounting assigns zero API cost to locally deployable models, which makes the reported savings depend partly on that convention. The detector’s training labels are also retrospective annotations of trajectories, so real-world performance may differ when tasks, tools, or model behavior shift.

For developers, the main lesson is to monitor task state, not just prompt difficulty or turn count. A useful production policy could begin with a stronger model, watch for an implemented and tested artifact, and hand off when remaining work becomes localized. Teams should validate this policy against their own workloads, measure recovery failures, and choose context transfer separately for each model pair.

AgentBaton does not show that smaller models can replace stronger ones. It suggests a narrower and more actionable possibility: expensive capability may be most valuable during the uncertain phase of an agentic task, while cheaper computation can take over after the path to completion has been established.


Sources

AI-assisted research summary. Check the original paper for complete methods, results, and limitations.

Avatar photo

Tech News

Jeff is a passionate blog writer who shares clear, practical insights on technology, digital trends and AI industries. With a focus on simplicity and real-world experience, his writing helps readers understand complex topics in an accessible way. Through his blog, Jeff aims to inform, educate, and inspire curiosity, always valuing clarity, reliability, and continuous learning.

↑ ↓ navigate Enter open Esc close