If an AI agent has 2,209 possible skills, why should it read all of them before it even knows the job? That is the practical question behind SkillWeaver, a new Alibaba research framework that cuts agent token use by more than 99% by retrieving only the tools a workflow actually needs, according to VentureBeat.
The pitch is simple and sharp. Instead of dumping a full tool library into a model’s context window, SkillWeaver breaks the user request into subtasks, retrieves relevant tool candidates for each one, then builds an execution graph that wires those tools together.
That matters because agent costs don’t only come from model choice. They also come from context bloat. Related cost pressure is already visible in developer debates around Claude Sonnet 5 agent costs and broader concerns that AI token costs can strain cybersecurity budgets.
Why are bloated tool prompts breaking enterprise AI agents?
Enterprise agents are becoming tool-heavy. They may need to call APIs, database functions, finance tools, cloud infrastructure actions, or Model Context Protocol (MCP) skills. The larger the library, the harder routing becomes.
A naive approach gives the model every available tool name and description, then asks it to pick. That burns tokens fast. Worse, it can still fail. The Alibaba researchers found that simply exposing a large model to all available tools did not reliably produce the right tool category.
Real business workflows make the problem harder. A request like this is not a one-tool job:
"Download the dataset, transform it, and create visual reports"
That requires sequencing. First an API or fetch tool. Then a data transformation tool. Then a charting tool. If one output does not match the next input, the workflow breaks.
XOOMAR analysis: The important shift here is from “which tool should the model call?” to “what workflow should exist before any tool is called?” SkillWeaver treats routing as planning first, retrieval second, execution third. That is a better fit for production agents than one-shot tool selection.
What problem does SkillWeaver solve in large AI tool libraries?
A skill is a reusable tool specification described in structured natural language. It tells the agent what the tool does and how it fits into a workflow.
The hard part is matching messy user intent to the exact vocabulary of the skill library. A human may ask to “clean up the file.” The available skill may be documented as “csv-parser” or “etl-pipeline.” If the model decomposes the task too vaguely, retrieval misses.
Alibaba frames this as compositional skill routing. The agent must do three things at once:
- Decompose: Split the request into atomic subtasks.
- Retrieve: Match each subtask to candidate skills.
- Sequence: Combine the selected skills into a usable plan.
That is different from single-skill routing. Enterprise tasks are usually chains, not isolated calls.
The research points to a blunt bottleneck: decomposition quality matters more than just scaling the model. Bigger models can still choose badly if their plan does not match the available tool library.
How does SkillWeaver turn one business request into a tool execution graph?
SkillWeaver runs through three stages: Decompose, Retrieve, and Compose.
| Stage | What happens | Why it matters |
|---|---|---|
| Decompose | An LLM splits the user request into subtasks | Bad subtasks poison retrieval |
| Retrieve | An embedding model searches the skill library for candidates | The agent sees only relevant tools |
| Compose | A planner checks compatibility and builds a graph | Outputs must fit downstream inputs |
The final plan is represented as a Directed Acyclic Graph (DAG). In plain terms, it is a dependency map with no circular loops. It shows which tool must run before another and which steps might run in parallel.
Take the dataset example. SkillWeaver may split it into:
- Download the dataset.
- Transform the data.
- Create visual reports.
The retriever may find “api-client” or “http-fetch” for the first task, “csv-parser” or “etl-pipeline” for the second, and “chart-gen” for the third. The compose stage then selects the combination that works together, not just the best-looking tool for each isolated step.
Compatibility is the quiet killer here. The best retrieval result for step two is useless if it cannot consume the output from step one. SkillWeaver’s graph-based plan tries to catch that before execution.
How does Skill-Aware Decomposition stop agents from choosing the wrong skills?
The key addition is Skill-Aware Decomposition (SAD). It is not a one-shot plan. It is a feedback loop.
First, the LLM drafts a decomposition. Then the system retrieves loosely matching tools. Those tool candidates are fed back to the LLM as hints. The LLM rewrites the subtasks so their wording and granularity better match the actual skill library.
That sounds small. The benchmark says it is not.
In the vanilla setup, Qwen2.5-7B-Instruct predicted the correct number of steps 51.0% of the time. With SAD, that rose to 67.7%. With Qwen-Max, decomposition accuracy reached 92%. On hard tasks requiring four to five skills, SAD improved accuracy by 50%.
The model-size result is the warning shot. A larger 14-billion parameter model performed worse than the 7B model in the unguided vanilla setup because it over-decomposed tasks into tiny, unnecessary steps. Once SAD supplied retrieved tool hints, accuracy improved.
XOOMAR analysis: For teams building agents, this argues against reflexively buying bigger models to fix routing. The cheaper fix may be giving the model better evidence about the tools it actually has.
What did the SkillWeaver benchmark show about accuracy, cost, and failure modes?
Alibaba tested SkillWeaver on CompSkillBench, a custom benchmark with 300 multi-step queries. The skill library included 2,209 real-world skills from the public MCP set, spanning 24 functional categories including cloud infrastructure, finance, and databases.
The token result is the headline. The brute-force LLM-Direct baseline used an estimated 884,000 tokens per query. SkillWeaver used roughly 1,160 tokens per query. That is a 99.9% reduction.
The brute-force method also struggled. Even with strong task breakdown capabilities, Qwen-Max retrieved the right tool category only 21.1% of the time when flooded with tool options. The ReAct-style agent loop failed completely on decomposition accuracy, scoring 0%, because it collapsed multi-step plans into isolated actions.
Implementation is possible, but not turnkey. The researchers have not released SkillWeaver’s source code. They did share prompt templates in the paper, and the system uses off-the-shelf pieces: all-MiniLM-L6-v2, FAISS, and standard orchestration patterns. Swapping in BGE-base-en-v1.5 improved accuracy without fine-tuning.
There are still production gaps. The framework plans and routes, but it does not solve error recovery. If an API call fails in step two, the chain can break. Teams will still need retries, fallbacks, reranking, and validation around the graph.
The practical watch item is clear: if SkillWeaver-style routing works outside the benchmark, agent builders may spend less time expanding context windows and more time maintaining clean, searchable skill libraries. That is where the next cost fight in AI agents is likely to move.
The Bottom Line
- Tool-heavy enterprise agents can become expensive when every available skill is loaded into context.
- SkillWeaver targets context bloat by retrieving only the tools needed for each workflow step.
- Lower token use could make complex AI agents more practical for business workflows involving APIs, databases, and reporting tools.
Originally published on XOOMAR. For more news and analysis, visit XOOMAR.
Top comments (0)