Alibaba Develops Technology to Reduce AI Agent Token Consumption by 99%

China's tech giant Alibaba's research team has developed SkillWeaver, a framework that significantly improves the operational efficiency of AI agents. The framework adopts an approach of systematically decomposing complex tasks and selecting only necessary tools, achieving over 99% reduction in token consumption—the unit of characters and words that AI processes—compared to the conventional method of loading an entire tool library into the AI at once, while also improving accuracy, the research team reports.

The background to this research lies in the challenge of "tool explosion" accompanying the expansion of enterprise AI systems. In recent times, AI agents increasingly work in conjunction with large-scale systems equipped with hundreds of types of tools and functions (skills), and determining "which tool to use at which step" has become extremely difficult. Most current frameworks adopt a "one-shot selection" approach, passing the entire library to the model at once for selection; however, this quickly exceeds the context window (the amount of information that can be processed) and escalates costs. SkillWeaver is an attempt to directly address this problem.

The core of SkillWeaver lies in its approach of converting the entire given task into a structure called an "execution graph" and assigning optimal skills to each step. Furthermore, by combining with a technique called "Skill-Aware Decomposition (SAD)," it realizes a feedback loop in which the agent iteratively acquires, validates, and narrows down candidate tools. For example, in response to an instruction like "download a dataset, process it, and create a graph report," it automatically constructs a plan that allocates API clients, data processing, and visualization tools to separate stages and executes them sequentially.

In experiments, using this framework resulted in over 99% reduction in token consumption compared to indiscriminately loading all tools, while also improving accuracy. The research team also points out that the bottleneck in accuracy is not model performance but rather "the granularity of task decomposition." In other words, the key to correct tool selection lies in how appropriately a task can be divided into smaller units.

SkillWeaver has high compatibility with integration platforms for multiple tools, such as MCP (Model Context Protocol), and is envisioned for application in scenarios where a series of business operations—data acquisition, transformation, and report generation—are autonomously handled. MCP is a mechanism that enables AI agents to invoke external tools or services in a standardized manner and has been rapidly spreading in recent years.

In bringing AI agents to a practical level, optimizing tool selection is an important challenge in both cost and performance aspects. Reduction in token consumption directly leads to lower operational costs, making it a significant advance for enterprises handling large-scale agent systems. Design philosophies like SkillWeaver's approach of "retrieving only what is needed, when it is needed" have the potential to influence the future direction of agent development. The trajectory of research commercialization and external disclosure will remain a focal point of attention.