AI Agents Master Complex Tasks with SkillWeaver

Navigating complex workflows in enterprise AI requires accurately routing subtasks to the correct tools and skills. Alibaba researchers have developed SkillWeaver, a framework designed to tackle this challenge by constructing an execution graph for tasks and selecting the optimal skills for each step. A key innovation is Skill-Aware Decomposition (SAD), which employs an iterative feedback loop to allow agents to fetch and verify potential tool candidates.

This method contrasts with existing frameworks that perform tool selection in a single step. SkillWeaver orchestrates task handling through three stages: Decompose, Retrieve, and Compose. Initially, an LLM breaks down complex queries into atomic sub-tasks. Next, an embedding model identifies a shortlist of candidate tools for each sub-task from a vast library.

Finally, a planner evaluates these candidates for compatibility, ensuring seamless integration between tools and composing them into a Directed Acyclic Graph (DAG) for execution. Experiments using the CompSkillBench benchmark with a library of 2,209 real-world skills demonstrated SkillWeaver's effectiveness. The SAD feedback loop improved decomposition accuracy significantly, with accuracy reaching 67.7% for a 7B model and 92% for a larger model.

SkillWeaver achieved a 99.9% reduction in token consumption compared to naive methods, reducing usage from an estimated 884,000 tokens to approximately 1,160 tokens per query. This translates to substantial cost savings and faster response times for AI practitioners. While the source code is not yet public, the framework's components are reproducible using off-the-shelf tools and prompt engineering techniques.