CRAFT post-training cuts schema-stuffed agent tokens while lifting analytics scores

Researchers describe CRAFT, a two-stage post-training recipe for schema-grounded enterprise coding agents, in an arXiv cs.AI preprint. The work targets agents that turn natural-language analytics requests into code over proprietary APIs, schemas, and metric definitions, arguing that stuffing full schema and tool docs into every prompt raises inference cost and hurts multi-turn reliability. Stage one uses schema-stripped PLAN supervised fine-tuning so agents learn domain-structured plans and executable behaviors from validated trajectories without exhaustive prompt-time schema injection. Stage two applies execution-shaped reinforcement learning for tool selection, code quality, plan-code consistency, and recovery from failed runs. Trajectories pass a Tri-Gate filter of execution validation, data-integrity checks, and an LLM-judge reasoning audit. The authors evaluate CRAFT for planned rollout in advertising analytics across campaign analysis, metric drill-downs, entity-level work, and multi-turn refinement. Their enterprise test bed uses beta APIs, 25 schema-linked core entities, and 30 agentic workflows. Versus a schema-stuffed baseline, CRAFT improved composite Agent Score by +9.6 pp, consistency by +4.1 pp, and multi-turn coherence by +4.2 pp, while cutting input-token burden by about 9x and schema-discovery loops by up to 5x.