WEFT is a framework for scaling tool-use post-training of general-purpose AI agents by jointly coordinating environments, tasks, agent harnesses, and evaluators rather than scaling environments alone. It introduces execution-driven self-evolution, which analyzes failure traces to iteratively refine these components, combined with prefix-preserving sampling and atomic-turn credit assignment to stabilize large-scale training, plus a system called MegaMCP that manages concurrent rollouts over shared tool services. In evaluations, the resulting WEFT-14B model outperformed Agent-World-14B by 2.23 to 12.27 percentage points across benchmarks including BFCL V4 and Toolathlon-Verified.
