New Alibaba AI framework skips loading every tool, cutting agent token use 99%

Alibaba researchers introduced SkillWeaver, a framework that builds execution graphs for complex tasks and routes each step to the most relevant skills. Its Skill-Aware Decomposition technique iteratively retrieves and vets tool candidates, improving accuracy while reducing token consumption by more than 99% versus exposing agents to full tool libraries.

The work highlights tool routing as a growing bottleneck for enterprise AI agents as they connect to larger ecosystems of APIs, skills, and protocols. More efficient decomposition and retrieval methods could make multi-step agents cheaper, more reliable, and easier to deploy in real business workflows.