Appier introduces SMITH framework to enable AI agents to build and refine their own tools

Appier has announced a new reinforcement learning framework called SMITH (Schema-grounded Multi-task Iterative Tool Honing), which has been accepted for presentation at the NeurIPS artificial intelligence conference.

The SMITH framework addresses a common limitation in Agentic AI where models can create tools but struggle to utilize them effectively. By integrating tool creation and usage into a single training loop, the framework allows AI agents to build, test, and continuously refine tools based on real-world problem-solving results.

Research indicates that smaller models trained with SMITH can develop reusable tools that rival the performance of much larger models on unseen tasks. Furthermore, these tools can be shared across different models and tasks, which significantly reduces token costs by replacing repetitive reasoning with direct tool calls. In experimental settings, the framework helped reduce average output token counts from 3,206 to approximately 100, representing a 32-fold increase in efficiency.