Alibaba researchers unveiled SkillWeaver, an AI framework that cuts token consumption by 99 percent when routing enterprise agents to the right tools. The system solves a critical bottleneck in complex AI workflows where agents juggle hundreds of tools and often misidentify which one handles each task step.

SkillWeaver builds an execution graph for a given task, then assigns the optimal skill to each node rather than loading all available tools upfront. This selective loading dramatically reduces computational overhead. The framework pairs this with Skill-Aware Decomposition (SAD), a feedback loop technique that lets agents iteratively fetch and validate relevant tool candidates instead of evaluating the entire tool library at once.

The compositional approach marks a departure from existing tool-routing frameworks. Rather than choosing tools blindly or loading every option into context, SkillWeaver uses dynamic feedback to surface only the most relevant candidates. For enterprises deploying AI agents across customer service, supply chain, and internal operations, the 99 percent token reduction translates directly to faster inference, lower API costs, and better scalability.

The paper, posted on arXiv, arrives as enterprises demand more sophisticated AI systems capable of handling multi-step workflows. Companies like ServiceTitan, HubSpot, and Zapier have already built agent platforms that route tasks across dozens of integrations. SkillWeaver addresses the same problem from the model layer, enabling cheaper and faster execution without requiring enterprises to redesign their agent architectures.

Token efficiency matters enormously in production deployments. Each percentage point of reduction compounds across thousands of daily requests, affecting both margins and user experience. Alibaba's work shows that intelligent graph construction and iterative tool selection beat the brute-force approach of loading everything.

The framework remains research-stage, but the implications ripple across the enterprise AI market. Alibaba's contribution gives open-source projects