CRAFT Boosts Enterprise Coding Agents, Reduces Schema Burden.

Enterprise coding agents are designed to convert natural language analytical requests into executable code, interacting with proprietary APIs, schemas, and metric definitions. A common issue is that injecting exhaustive schema documentation into every prompt increases inference overhead, complicates schema evolution, and reduces reliability in multi-turn analyses. Researchers introduce CRAFT, a two-stage post-training approach that enables agents to acquire stable schema knowledge and tool-use behavior without constant prompt-time schema injection. The first stage, "schema-stripped PLAN supervised fine-tuning," teaches domain-structured plans and executable behaviors from validated trajectories. The second stage uses "execution-shaped reinforcement learning" to align the agent's policy for tool selection, code quality, plan-code consistency, and recovery from execution failures. Training trajectories are meticulously curated using a Tri-Gate filter for validation. Evaluated in advertising analytics, CRAFT improved composite Agent Score by +9.6 pp, consistency by +4.1 pp, and multi-turn coherence by +4.2 pp, while drastically reducing input token burden by 9x and schema-discovery loops by up to 5x compared to a schema-stuffed baseline.