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 capacity to use them reliably. Existing systems often rely on manual engineering or separate models for creation and execution, which prevents effective feedback. SMITH resolves this by treating tool descriptions and parameter specifications as direct training signals. When a tool fails or provides vague outputs, the system updates its logic, fostering a continuous cycle of improvement.

Experimental results demonstrate that a 4-billion-parameter model utilizing SMITH can generate tools that exceed the performance of 30-billion-parameter systems. These tools are not only reusable but also interoperable across varying model sizes, including lightweight variants as small as 350 million parameters. By converting complex, repeated reasoning into persistent tools, the framework reduced output from 3,206 tokens to approximately 100 tokens per task. This shift toward reusable, verified capabilities offers a path for enterprises to automate repetitive workflows—such as financial reporting or customer service routing—without the overhead of constant re-computation.