Enterprise that turn natural language into code usually inject full schema documentation into every prompt. This approach is costly, fragile to schema changes, and hurts multi-turn reliability. CRAFT instead embeds stable schema knowledge directly into the model through a two-stage post-training recipe, eliminating the need for exhaustive prompt-time schema injection.
CRAFT first uses schema-stripped supervised fine-tuning (PLAN SFT) to learn domain plans and executable behaviors from validated trajectories. Then execution-shaped reinforcement learning (GRPO) aligns the policy for tool selection, code quality, and recovery from failures. Training data is curated by a Tri-Gate filter combining execution validation, data-integrity checks, and an LLM-judge reasoning audit.
Evaluated in advertising analytics over 25 schema-linked entities and 30 workflows, CRAFT improved composite Agent Score by 9.6 percentage points, consistency by 4.1 pp, and multi-turn coherence by 4.2 pp compared to a schema-stuffed baseline. It also reduced input-token burden by roughly 9x and schema-discovery loops by up to 5x. These results show that post-training can replace prompt stuffing for enterprise coding agents, lowering cost and improving maintainability.