Appier Unveils SMITH Framework to Automate AI Tool Creation

The SMITH (Schema-grounded Multi-task Iterative Tool Honing) framework addresses a persistent bottleneck in Agentic AI: the disconnect between an AI’s ability to generate tools and its proficiency in using them. Traditional systems often rely on engineers to manually configure APIs, or they separate the creation and execution phases, which limits the model’s ability to learn from real-world performance. SMITH closes this loop by feeding feedback from tool failures or parameter errors directly back into the training process.

Experimental results demonstrate that a 4-billion-parameter model utilizing SMITH can generate tools that surpass those created by 30-billion-parameter models. These tools are inherently reusable, allowing models as small as 350 million parameters to tackle complex, unseen tasks. Beyond performance, the framework offers significant computational efficiency. By converting repetitive reasoning processes into stable, callable tools, the system reduced output requirements from over 3,200 tokens to approximately 100, marking a 32-fold increase in processing speed.

According to Appier CEO Dr. Chih-Han Yu, this evolution mirrors human learning, where past problem-solving experiences are codified into tools to avoid redundant effort. For enterprise applications—ranging from financial data processing to customer service automation—this means agents can now build a shared library of proven methods. As these agents encounter new tasks, they can leverage existing assets rather than regenerating logic from scratch, effectively turning historical data into a scalable infrastructure for business operations.