Key Takeaways

  • Token Efficiency: Consumption drops from 884,000 to just 1,160 tokens per query, alongside a 50% increase in accuracy on complex tasks.
  • Technology: The SkillWeaver framework runs on Skill-Aware Decomposition (SAD), built on top of the 7-billion-parameter Qwen2.5 model.
  • Operational Limit: The planner doesn't handle runtime errors; a single API failure mid-chain halts the entire execution process.

The Information Noise Problem

When a language model has to orchestrate thousands of external tools, brute force doesn't work. Dumping the entire library of available functions into a single prompt creates an unsustainable load: lab tests hit peaks of 884,000 tokens just to let the agent "see" its available options. It's a bottleneck that SkillWeaver promises to eliminate entirely, shifting the approach from static to dynamic.



SkillWeaver: -99% Tokens, +50% Accuracy in AI Agents - Foto 1

SkillWeaver: -99% Tokens, +50% Accuracy in AI Agents - Foto 2

The Logic of Decomposition

The core of the system is Skill-Aware Decomposition. Instead of loading everything at once, the algorithm breaks the user's request into a node-based execution graph. For each individual sub-goal, it retrieves only the genuinely relevant tools, filtering out background noise. Test results, run on a proprietary benchmark of 300 multi-step requests drawing from 2,209 real tools across the MCP ecosystem, speak for themselves: consumption cut to 1,160 tokens with a 50% jump in accuracy. One notable technical finding stands out: large models, when left without structural guidance like SAD, tend to fragment tasks into microscopic, unnecessary steps, degrading final output instead of improving it.

Closed Code, Open Templates

Developed in Shanghai labs, the project hasn't yet released its full source code. However, the prompt templates used are already public and can be reproduced using standard libraries like LangChain or custom Python scripts. A notable structural gap remains: runtime error handling is missing from the planner. If an API call fails during an intermediate step, the entire logical chain breaks down with no automatic recovery protocol in place.