AI systems managing complex business tasks often struggle because they must select the correct skill from hundreds of available options, which presents a major challenge for developers. This difficulty means that agents frequently get confused when they attempt to split a large request into smaller pieces. Alibaba researchers recently developed [SkillWeaver](https://github.com/OSU-NLP-Group/SkillWeaver), a novel AI framework designed to solve this difficult routing problem by creating a structured execution graph for any task. This system uses Skill-Aware Decomposition (SAD), which allows the agent to find and check relevant tools repeatedly instead of selecting them all at once. This compositional approach and the feedback loop mechanism distinguish SkillWeaver from other frameworks that attempt to choose tools in a single, one-shot effort. SkillWeaver relates to real-world uses where autonomous AI agents manage ecosystems of multiple tools, such as the Model Context Protocol, to run multi-step business operations. These operations might involve downloading data, changing information, or generating visual reports, requiring many interconnected skills.
**The Challenge of Skill Routing**
Skills represent a core pattern in modern large language model agent designs, acting as modular and reusable tool specifications. As enterprise agents integrate with massive libraries of tools, correctly guiding user questions to the proper skills becomes an inefficient and difficult task. Exposing an entire library of skills to a large language model quickly overwhelms the context limits and consumes a massive amount of processing tokens. Most current tool-use systems try to solve this problem by using API retrieval or matching documentation, but these methods fail for complex tasks. These current frameworks often treat routing as a single-skill selection problem, which does not accurately reflect real-world, multi-part requests. For example, a standard business request, such as downloading data and creating reports, cannot be completed by just one tool. It demands breaking the prompt down and sequencing an API client, a data processor, and a visualization tool into a coherent, multi-step plan. The difficulty lies in the fact that real-world queries inherently consist of multiple actions, requiring a dynamic approach rather than just single-skill selection. SkillWeaver frames the problem of handling complex tasks that require many skills as ‘compositional skill routing,’ demanding a new level of planning. Given a huge list of tools, an agent must simultaneously figure out how to split the request into atomic sub-tasks and map each sub-task to the best available skill. SkillWeaver orchestrates this complex process through three distinct stages: Decompose, Retrieve, and Compose.
**How SkillWeaver Works**
In the first stage, the large language model acts as a task splitter, breaking the user’s complex query into a sequence of sub-tasks. Once the sub-tasks are clearly defined, the [system](https://thedecisionlab.com/reference-guide/philosophy/system-1-and-system-2-thinking) uses an embedding model to compare each sub-task against the skill library, pulling a shortlist of top candidate tools. In the final stage, a planner evaluates the retrieved candidates based on how well they work together, checking for compatibility between skills. It then generates a final execution plan, which maps out required connections so that independent tasks can run in parallel.
* The model drafts an initial plan.
* The system searches for loosely matching [skills](https://arxiv.org/html/2603.12056v2) to use as hints.
* The model rewrites its plan to perfectly match the existing tools.
For instance, if a user asks an AI agent to download data and create reports, the splitter breaks this into three distinct sub-tasks. During the retrieve stage, the system finds candidates like ‘api-client’ or ‘csv-parser’ for the different steps in the workflow. Finally, the compose stage selects the specific combination of tools, like ‘api-client’ and ‘chart-gen,’ and wires them into an executable plan. **How Does SkillWeaver Improve Planning?**
A major challenge in these systems is that large language models often produce generic step descriptions that do not match the specific technical vocabulary of the actual available tools. To fix this, SkillWeaver introduces Iterative Skill-Aware Decomposition (SAD), which acts as a novel feedback loop within the system. SAD works by having the model draft an initial plan, then searching for loosely matching skills to use as hints for the model. It feeds these retrieved skills back into the LLM, allowing the model to rewrite its splitting to perfectly match the existing tools. This iterative refinement helps ensure the agent uses the right language to command the specific tools in the ecosystem. This feedback loop anchors the model to reality, ensuring that its abstract planning aligns with the concrete capabilities of the available skills.
**Performance in Real-World Scenarios**
The researchers built a custom benchmark called CompSkillBench to test SkillWeaver in realistic enterprise scenarios. This benchmark includes 300 multi-step queries of varying difficulty levels and uses a library of 2,209 real-world skills. These skills cover 24 different functional categories, such as cloud infrastructure, finance, and databases, to mirror complex business needs. For the core engine, the researchers primarily used a lightweight 7-billion parameter model for task splitting, paired with a semantic search retriever. SkillWeaver was tested against three setups: a brute-force ‘LLM-Direct’ method, a vanilla LLM splitting without SAD, and a standard ReAct-style agent loop. For related coverage, see [Amazon Framework Details Engineering Trustworthy Ai Agents For Enterprise Use](https://www.techmogo.com/ai/amazon-framework-details-engineering-trustworthy-ai-agents-for-enterprise-use/).
