OODA-Tool is a policy for that breaks each decision into four typed stages: Observe reconstructs the task state, Orient checks whether execution is warranted, Decide forms an admissible action structure, and Act grounds the final schema-valid call. A central controller checks each handoff, preventing actions from being generated directly from raw history. It is compared against direct function-calling and ReAct-style policies using Qwen3 models from 0.6B to 14B.
The approach improves task success across all model sizes, with larger gains on smaller models and on tasks where actions depend on information accumulated over turns and prior tool results. Benefits are smaller for larger models or tasks that mostly require many parallel calls. Direct generation remains cheaper when one-pass simplicity is the priority.