EnterpriseBench Corecraft: Training Generalizable Agents on High-Fidelity RL Environments

Surge AI introduces CoreCraft, the first environment in EnterpriseBench for training AI agents on realistic enterprise workflows. Training GLM 4.6 on this high-fidelity customer support simulation improved task performance from 25% to 37% and showed positive transfer to other benchmarks, demonstrating that quality training environments enable generalizable AI capabilities.

• →CoreCraft simulates a complete customer support organization with 2,500+ entities and 23 tools to test AI agent capabilities on real-world tasks.
• →Leading frontier models like GPT-5.2 and Claude Opus 4.6 solve fewer than 30% of expert-evaluated tasks.
• →Training GLM 4.6 with GRPO improved performance from 25.37% to 36.76% on held-out evaluation tasks.
• →The training gains transferred to other benchmarks with improvements of 4.5-7.4% across different domains.
• →High-fidelity, diverse training environments with expert evaluation are key to developing generalizable AI agents.