PaperlayerView exact revision on arXiv

Source-linked AI summary

AgentBaton: Efficient Model Collaboration in Agentic Tasks

Junlong Li, Jindong Jiang, Yonggan Fu, Jiarui Yao, Hanrong Ye, Wonmin Byeon, Peter Belcak, Hao Zhang, Pavlo Molchanov

arXiv:2610.35959v1cs.MAcs.AI

TL;DR

Long-horizon agents must balance the high cost of capable models against the lower cost but often lower accuracy of efficient models, and query-level routing does not address capabilities that evolve during a task. The paper studies one-time model handoffs and introduces AgentBaton to detect suitable capable-to-efficient handoff points. Across benchmarks, it reports lower API costs while largely preserving accuracy, and an extended empirical accuracy–cost Pareto frontier across reasoning-effort settings.

  • Problem

    Whether capable and efficient models can collaborate within a long-horizon task to reduce cost while preserving accuracy is less explored than query-level routing, despite task requirements evolving across turns.

    Source: §1 · p. 2

  • Method

    The paper compares handoff directions and history-transfer methods, then uses AgentBaton to detect suitable capable-to-efficient handoff points online from recurring execution patterns.

    Source: §1 · p. 2

  • Results

    Across 11 model pairs and two benchmarks, AgentBaton reduces API costs by 65.9% on average while largely preserving accuracy, with gains up to 8.8 percentage points over capable-only execution.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The experiments indicate that task progress can serve as a signal for dynamic capability allocation, and handoffs can extend the empirical Pareto frontier beyond single-model configurations across reasoning-effort settings.

    Source: §6 · p. 10

Abstract

from arXiv · show

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.

1. Introduction

Long-horizon agents face a cost–accuracy trade-off between capable models and cheaper, often less accurate efficient models, while query-level routing does not account for capabilities changing during a task. The paper investigates one-time handoffs and introduces AgentBaton, a detector for suitable handoff points.

1. Introduction

  • The paper studies handoff direction and two history-transfer methods: placing prior interaction history in context or making it available through an inspectable logfile.In either case, the receiving model continues from the predecessor’s workspace.Source: §1 · p. 2
  • Capable-to-efficient handoffs at suitable task stages can largely preserve accuracy while reducing cost, whereas efficient-to-capable handoffs generally underperform capable-only execution and may offer little or no savings.The authors suggest established progress from a capable model can enable an efficient model to finish work, while frequent switching risks accumulating trajectory errors.Source: §1 · p. 2
  • Across 11 model pairs and two benchmarks, AgentBaton cuts API costs by 65.9% on average while largely preserving accuracy, with gains up to 8.8 percentage points over capable-only execution.The paper also reports that AgentBaton extends the empirical accuracy–cost Pareto frontier beyond evaluated single-model configurations across reasoning-effort settings.Source: §1 · pp. 2–3
  • AgentBaton uses recurring patterns in execution history to detect suitable capable-to-efficient handoff points online, after which the efficient model continues the task.The capable model starts the task; the detector evaluates progress after each turn.Source: §1 · p. 2

2. Experimental Design

The experimental design studies one-time handoffs, with the receiving model inheriting the source model’s workspace and continuing to termination. It compares context-transfer methods across two benchmarks and reports accuracy and API cost at selected checkpoints.

2. Experimental Design

  • A one-time handoff transfers control after several turns; the receiver continues to termination, inherits workspace state, and incurs cost alongside the source model’s pre-handoff cost.Execution history is transferred either directly in context or through a chronological JSONL logfile that the receiver can inspect.Source: §2 · p. 3
  • The evaluation uses Terminal-Bench 2.1 and DeepSWE v1.1, with a shell-based mini-SWE-agent scaffold, a two-hour task timeout, and up to 16K newly generated tokens per turn.Terminal-Bench spans terminal-based workflows; DeepSWE focuses on long-horizon software engineering tasks in active repositories.Source: §2 · pp. 3–4
  • Table 1 reports accuracy and mean recorded API cost per task for transfer methods at 30% and 75% of the source model’s baseline trajectory.Bold marks strictly higher accuracy or lower cost between methods at the same checkpoint.Source: Table 1 · p. 4

3. Understanding Model Handoffs in Agentic Tasks

The experiments find that handoff direction and timing matter: suitable capable-to-efficient handoffs can preserve accuracy with substantial savings, while efficient-to-capable handoffs do not consistently achieve both. Successful handoff readiness is associated with established work, task-relevant feedback, and more focused remaining tasks.

3.1. Handoff Direction and Context Transfer

  • Transfer preference depends on direction and model: logfile transfer generally matches or exceeds in-context accuracy and lowers cost for efficient-to-capable handoffs, while the preferred format varies for capable-to-efficient handoffs.For capable-to-efficient handoffs, in-context transfer is more effective for GLM-5.3 and Kimi K3 in most cases; GPT-5.6 Sol results are mixed across checkpoints and benchmarks.Source: §3.1 · p. 4

3.2. When Does Handoff Preserve Accuracy?

  • With suitable timing, capable-to-efficient handoffs approach capable-only accuracy while reducing cost: Kimi K3 at 45% saves 61.5%, GPT-5.6 Sol at 60% saves 47.9%, and GLM-5.3 at 30% saves 82.4%.The cited examples report respective accuracies of 72.57% versus 73.45%, 67.26% versus 68.10%, and 61.1% matching its capable-only baseline.Source: Figure 3 · p. 5, §3.2 · p. 5
  • Efficient-to-capable handoffs do not consistently preserve accuracy with substantial savings; at 30% on DeepSWE, GPT-5.6 Sol reaches 70.80% accuracy versus 68.10% alone, with only 6.1% lower cost.Later handoffs to GLM-5.3 and Kimi K3 generally reduce cost but also lose accuracy.Source: Figure 3 · p. 5, §3.2 · p. 5

3.3. Patterns of Successful Handoffs

  • Figure 4 contrasts earlier failed Qwen continuations with later successful ones after intervening capable-model progress, across illustrative cases from two benchmarks and two capable models.The cases include fixing a quoted-path issue, validating a coordinate transformation, and localizing an encoding fault.Source: §3.3 · p. 5, Figure 4 · p. 6
  • Readiness signs include a substantive implementation or artifact, task-relevant feedback from execution or inspection, and remaining work narrowed to diagnosis, repair, validation, or finalization.Feedback can be informative even when an execution fails, if it narrows the unresolved problem.Source: §3.3 · p. 6

4. AgentBaton

AgentBaton uses a lightweight detector to identify capable-to-efficient handoff points online. Across the evaluated settings, it reduces API costs while often preserving accuracy, outperforming routing baselines and extending the empirical accuracy–cost frontier.

4.2. Training AgentBaton

  • AgentBaton’s online detection is described with separate, growing message histories; the passages motivate a lightweight detector to recognize handoff points during capable-model execution and enable transfer when observed progress indicates readiness.AgentBaton achieves a median latency of 338 ms and a 95th-percentile latency of 1.11 s on a single NVIDIA H100 GPU.Source: §4 · p. 6, Figure 5 · p. 7, §4.2 · p. 7

4.1. Prompt-based handoff

  • Prompted handoffs varied by model, and consistent handoff behavior did not guarantee a favorable accuracy–cost trade-off.Kimi K3 never handed off, GLM-5.3 did so on only 6 of 113 tasks, and GPT-5.6 Sol handed off on every task but still cost $3.75 per task versus $5.32 for capable-only execution.Source: §4.1 · pp. 6–7, Table 2 · p. 6

4.3. Experimental Results

  • On DeepSWE, 12 of 22 model-pair settings improved accuracy or incurred a relative loss below 5%; Kimi-to-Qwen achieved 72.6% accuracy versus 73.5% capable-only while reducing cost by 72.9%.The table reports results across model pairs and benchmarks; GPT-5.6 Luna also yielded accuracy above capable-only for four of five capable models on DeepSWE.Source: Table 3 · p. 8, §4.3 · p. 8

4.4. Comparison with Routing and Handoff Baselines

  • AgentBaton reduced API costs by 46.7–86.6% relative to MTRouter while matching or improving accuracy in all four settings.A query-level router trained on initial task descriptions selected the efficient model for only one task, effectively yielding capable-only execution with little cost savings.Source: §4.4 · pp. 8–9
  • AgentBaton achieved accuracy close to a fixed-progress baseline despite that baseline’s access to future trajectory length, and it consistently outperformed fixed-turn handoffs in accuracy.AgentBaton incurred equal or lower API cost than the fixed-progress baseline.Source: §4.4 · p. 9, Table 4 · p. 9

4.5. Pushing the Accuracy–Cost Pareto Frontier

  • Across the evaluated reasoning-effort settings, AgentBaton extended the empirical accuracy–cost frontier beyond single-model configurations.For example, Kimi-to-Luna with high effort reached 79.6% accuracy at $1.60 per task, above the 75.2% single-model maximum at $8.90.Source: §4.5 · p. 9

5. Related work

Prior work routes individual requests, orchestrates models across turns, or switches models within reasoning outputs. AgentBaton instead learns when to hand off between complete agent turns, given a capable–efficient model pair.

  • Request-level routers and cascades select models for individual queries, but interactive agents change the workspace and inherited state as they act.Examples include FrugalGPT [7], RouteLLM, RouterDC [9], and Route to Reason [27].Source: §5 · p. 9
  • Multi-round orchestration learns model or tool coordination from execution feedback, while AgentBaton separates handoff timing from model selection for a given model pair.MTRouter [25] performs turn-level routing; AgentBaton’s detector assesses progress after each turn and the receiving model continues until termination.Source: §5 · pp. 9–10
  • Prior handoff studies examine transitions using risk signals or predefined switch points, whereas AgentBaton learns capable-to-efficient handoff points online from trajectory patterns.SWE-Router [32], TACIT-Switch [33], and Handoff Tax [34] study model transitions and trajectory transfer.Source: §5 · p. 10
  • Reasoning-output methods transfer segments within one reasoning trajectory; AgentBaton transfers between complete agent turns and carries forward workspace state and interaction history.The receiving model continues with tools and persistent task-state changes until termination.Source: §5 · p. 10

6. Conclusion

The paper studies one-time model handoffs within long-horizon agentic tasks and introduces AgentBaton to identify suitable capable-to-efficient transfer points. Across two benchmarks, the approach reduces API costs while largely preserving or sometimes improving accuracy, and can extend the empirical Pareto frontier.

  • Capable-to-efficient handoffs offer favorable accuracy–cost trade-offs, especially after implementation, feedback, and remaining work have narrowed to localized refinement.AgentBaton dynamically identifies suitable handoff points; across DeepSWE v1.1 and Terminal-Bench 2.1, it substantially reduces API cost while largely preserving or occasionally improving accuracy.Source: §6 · p. 10
  • Reasoning-effort sweeps show that handoffs can extend the empirical Pareto frontier beyond single-model configurations.The conclusion identifies task progress as a signal for dynamic capability allocation.Source: §6 · p. 10

A. Handoff progress on Terminal-Bench

On Terminal-Bench 2.1, capable-to-efficient handoffs can improve accuracy and reduce cost at some checkpoints, but results vary across models and checkpoints.

  • At a 75% handoff checkpoint, Kimi K3-to-efficient achieved 87.64% accuracy versus 86.52% for Kimi alone, with 29.4% lower cost.Results were less consistent across models and checkpoints than on DeepSWE, and later handoffs did not monotonically improve accuracy.Source: A. Handoff progress on Terminal-Bench · p. 14

B. Handoff-point annotation protocol

The annotation protocol labels the earliest trajectory turn where readiness is established, using task descriptions and tool interactions rather than private reasoning or outcome information. It also records uncertainty, indexing, evidence, and provenance details.

B. Handoff-point annotation protocol

  • The annotator uses the task description and chronological tool calls and responses, excluding private reasoning, verifier outputs, final rewards, and continuation outcomes.Codex reviews trajectories retrospectively to locate the earliest suitable handoff point.Source: B. Handoff-point annotation protocol · p. 14
  • Readiness requires a substantive task artifact, observed task-relevant feedback, and remaining work focused on localized diagnosis, repair, validation, or finalization.Failed checks can qualify when they expose a localized problem; all tests need not pass, but unrelated checks or substitutes bypassing requested behavior are insufficient.Source: B. Handoff-point annotation protocol · pp. 14–15
  • The label marks the first jointly qualifying turn, even if later work regresses or substantially rewrites the implementation; trajectories with no such point receive no_clear_handoff_point.Readable failed or interrupted trajectories remain eligible, while missing or unreadable trajectories are recorded separately.Source: B. Handoff-point annotation protocol · p. 15
  • Turn and evidence indices count zero-based tool-bearing assistant steps, and each annotation record preserves traceability, readiness, supporting evidence, rationale, confidence, and caveats.For trajectories without a clear handoff point, handoff fields are null.Source: B. Handoff-point annotation protocol · p. 15

C. Complete AgentBaton results

Table 5 reports both context-transfer methods across the capable and efficient model combinations, while the main-text table selects a transfer method by source model and reports single-model baselines separately.

C. Complete AgentBaton results

  • Table 5 reports in-context and logfile transfer for all five capable models and both efficient models, plus GPT-5.6 Luna handing off to Qwen 3.8 27B; GPT-5.6 Sol and Luna use logfile transfer, other source models in-context.Source: C. Complete AgentBaton results · p. 15
  • The main-text table uses logfile transfer for GPT-5.6 Sol and GPT-5.6 Luna sources, and in-context transfer for the other source models.The table reports the selected transfer method for each source-model grouping.Source: C. Complete AgentBaton results · p. 15
  • Single-model baselines are reported separately in Table 3.Source: C. Complete AgentBaton results · p. 15

D. Complete reasoning-effort sweep results

The reasoning-effort sweep is documented through the numerical results table and the configurations plotted in Figure 1, including dominated configurations.

D. Complete reasoning-effort sweep results

  • Table 6 lists all single-model and AgentBaton configurations plotted in Figure 1, including dominated configurations.The Qwen baseline uses a 1M-token context window.Source: D. Complete reasoning-effort sweep… · p. 15
  • Table 6 uses Qwen and Luna as shorthand for Qwen 3.8 27B and GPT-5.6 Luna, respectively.Source: D. Complete reasoning-effort sweep… · p. 15
  • Table 6 reports the sweep's source-model effort configurations; accuracy is a percentage and cost is mean API cost in USD per task, with dashes marking unevaluated configurations.Source: Table 6 · p. 16
Loading 2610.35959v1…