Agentic AI
New Method Separates AI Reasoning from Action to Improve Multi-Step Tool Use
Researchers propose OODA-Tool, a framework that keeps AI agents focused on their current task state while generating reliable action sequences across multiple turns.
When AI agents must use tools across multiple turns—like calling functions, retrieving data, and acting on results—they often lose track of earlier information or make inconsistent decisions. A new approach called OODA-Tool aims to solve this by enforcing a clearer separation between understanding the current situation and deciding what to do next.
The Problem with Current Methods#
Today's AI agents typically rely on two dominant strategies: direct function-calling, which generates tool invocations one at a time, and ReAct-style policies, which interleave reasoning with action in a single forward pass. Both approaches train the model to handle state tracking and action generation simultaneously, creating what researchers call "state-action competition"—the pressure to output the next action can actually cause the model to forget or ignore information it gathered earlier in the conversation.
This problem grows worse in complex scenarios: multi-turn interactions where decisions depend on earlier tool results, situations with incomplete information, or tasks involving many tools. Smaller models struggle most.
How OODA-Tool Works#
The new framework adapts Boyd's Observe-Orient-Decide-Act cycle—a decision-making model from military strategy—to create a four-stage loop:
- Observe: The model reconstructs the full current task state from the interaction history.
- Orient: It checks whether taking an action is actually warranted at this moment.
- Decide: It forms a valid action structure that fits the constraints.
- Act: It generates the final external output.
Each stage acts as a checkpoint, ensuring that later decisions remain grounded in the state built during observation. This typed, closed-loop design prevents the agent from drifting away from what it actually knows.
Results Across Model Sizes#
Researchers tested OODA-Tool against conventional methods using Qwen3 models ranging from 0.6B to 14B parameters. The new approach showed consistent gains across all model sizes, with particularly strong improvements on smaller models and on tasks where decisions heavily depend on information accumulated over multiple turns or prior tool results.
Ablation studies—which test individual components of the framework—and transfer evaluations confirmed the robustness of the improvement. The framework generalizes well across different scenarios without requiring task-specific tuning.
