Tuesday 15 September 2026
Evening EditionRSS

ZOTPAPER

News without the noise

ZOTPAPER
Agent Harnesses

Separating state tracking from actions improves multi-turn tool reliability

OODA-Tool routes each decision through four controller-checked stages, improving task success across Qwen3 models from 0.6B to 14B parameters.

PaperIndependentcs.AIarXiv:2608.24368v1 ↗
Rongfeng Guo · Yinxuan Huang · Yusen Wu · Maoqing Zhong · Yunlu Chen · Meng Tang · +2 more
Research Digest··2 min read
Read the paper →
Thread:Agent Harness Optimization
Guo et al. introduce a structured tool-use policy that separates reconstructing task state, deciding whether to act, specifying an admissible action, and producing the final output. Across multi-turn, multi-tool, and incomplete-information evaluations, the authors report more reliable task completion than direct function calling and ReAct, especially for smaller models and tasks dependent on earlier context.

What they did

The authors designed OODA-Tool around four typed stages: Observe reconstructs the current task state; Orient determines whether execution is appropriate; Decide creates an admissible action structure; and Act produces the external call or response. A controller checks intermediate states before execution.

They compared this policy with direct function calling and ReAct-style reasoning using Qwen3 models ranging from 0.6B to 14B parameters. Evaluations covered multi-turn and multi-tool interactions, incomplete information, and tasks requiring agents to use details from earlier turns or previous tool outputs; controlled variants, ablations, and transfer tests examined which components contributed.

Key findings

  • OODA-Tool improved task success consistently across all tested Qwen3 model sizes relative to direct function calling and ReAct.
  • Gains were larger for smaller models, suggesting that explicit state-action separation can compensate for limited model capacity.
  • The strongest improvements appeared when correct actions depended heavily on information accumulated over several turns or returned by earlier tools.
  • Stage-level ablations and transfer evaluations supported the authors’ claim that the structured, controller-checked pipeline—not merely additional generation steps—drives the improvement.

Why it matters

The results suggest that reliable tool use need not depend solely on scaling the underlying model. Explicitly preserving task state and validating decisions before execution offers a practical form of agent scaffolding for workflows where an incorrect or premature call can derail subsequent steps.

Caveats

The supplied abstract does not report absolute success rates, benchmark names, latency, token overhead, or statistical uncertainty, so the magnitude and cost of the gains cannot be assessed here. The experiments are also limited to the Qwen3 model family, and broader validation is needed for other models, real-world APIs, longer interactions, and adversarial or rapidly changing environments.

§

Analysis

OODA-Tool fits a broader shift from unconstrained reasoning traces toward typed agent architectures with explicit state representations and execution gates. Its central hypothesis is that state tracking and action generation compete when learned within one autoregressive trajectory; the reported ablations provide evidence for that account, although comparisons controlling for inference-time compute remain important.

newspaper

Research Digest

Articles published under the Zotpaper byline are synthesized from multiple source publications by our AI editor and reviewed by our editorial process. Each story combines reporting from credible outlets to give readers a balanced, comprehensive view.

Editorial ProcessSubscribe