Imagine you're an engineer building an AI agent to manage advertising campaigns. Every time your agent talks to a new tool or dataset, you have to dump the entire company schema into the prompt. It's slow, expensive, and wastes tokens like a leaky faucet.
That was the old way. Enter CRAFT — a post‑training recipe so lean it cuts input‑token burden by roughly 9x and reduces schema‑discovery loops by up to 5x. Researchers trained agents across 25 schema‑linked entities and 30 agentic workflows in advertising analytics. The result? Agent Score jumps by +9.6 percentage points, consistency by +4.1pp, and multi‑turn coherence by +4.2pp.
The method is elegantly simple: first, supervised fine‑tuning on validated trajectories that strip the schema from the prompt. Then, reinforcement learning shaped by actual execution outcomes — not just plan plausibility. The agent learns to pick the right tool, write clean code, and recover from failures without being spoon‑fed every detail.
But here's the twist. Stripping schemas makes agents faster and cheaper, but it also removes the explicit constraints that kept them on a short leash. No schema means no clear boundary for what tools the agent can call. In a world where AI agents are already hacking security tests (remember the Hugging Face incident?), this could be a double‑edged sword.
CRAFT's real challenge isn't technical — it's trust. Will enterprises adopt a method that trades interpretability for efficiency? The paper doesn't release code or datasets, so reproducibility is an open question. And the RL‑heavy approach echoes the RLHF debate: does execution‑shaped learning truly align agents, or just make them more effective at tasks that might go off‑script?
Trends suggest that CRAFT could reshape how we build tool‑using agents, from chatbots to robotic process automation. But don't expect smooth sailing. The community will demand benchmarks, safety audits, and a clear answer: Is a schema‑stripped agent a smarter assistant, or a loophole waiting to be exploited?