Agentic Reasoning Compilation: From Traces to Executable Representations for Small Language Models

Language-model reasoning combines semantic interpretation with logical and numerical computation. Errors may therefore arise before computation begins: a model selects the wrong entities, omits a constraint, or confuses current state with history. We present agentic reasoning compilation, which combines external executors for supported computations with an orchestration pipeline guided by solution traces and execution feedback. The resulting system couples reusable typed representations and admission checks with deterministic execution, then uses planning, verification, and answer selection to handle remaining inputs. Reusable capabilities and control rules organize inference while language-model weights remain fixed. On the full BBEH suite, the system raises micro accuracy from 48.4% to 64.9% for gpt-5.6-luna and from 47.2% to 64.0% for qwen3.7-flash. Their adjusted harmonic means reach 47.7 and 47.9, exceeding gpt-5.6-sol's 37.2 at 67.2% and 97.7% lower estimated inference cost. A 100-item component comparison shows that orchestration preserves all 27 executor successes and solves 12 additional items. These results support a division of labor in which external execution supplies exact computation and orchestration extends reasoning coverage.