SPADE lifts tool-use agents 13.9 points

SPADE, a new self-play reinforcement-learning framework, reports a 13.9-point lift on ACEBench-Agent and a 5.7-point lift on BFCL-v4 multi-turn when a 30B model trains against LLM-generated, executable environments. Its environment designer writes state transitions, rewards, and verification code, then adapts tasks toward the reasoning agent’s capability frontier instead of recycling a fixed benchmark pool. The result is a research lead, not a deployment guarantee—the paper is marked work in progress—but it gives agent teams a sharper training direction: invest in executable task generation and accumulated environment memory, then verify gains on held-out work; the coordination study in this archive shows why adding agents without changing the protocol is not enough.