Appier introduces SMITH to train small AI agents to build and reuse tools
The AMW Read
Joint tool creation and use meaningfully advances small-model agent research with potential segment-wide efficiency benefits, but the reported benchmarks do not establish enterprise reliability or savings.
Appier introduces SMITH to train small AI agents to build and reuse tools
Taiwan-based Appier unveiled SMITH, a research method that jointly trains AI agents to create and use software tools. Developed with National Taiwan University researchers, the paper was accepted at NeurIPS, according to the report. The approach lets a model build a tool, test it and refine it within one learning loop. Tools enter a shared library only after passing tests on problems withheld from tool creation; improved versions can replace existing tools.
The market significance is the separation of tool-building capability from the model needed to execute recurring tasks. In the reported experiments, a 4-billion-parameter model trained across 13 reasoning tasks scored an average of 79.8 on unseen tasks, exceeding an untrained 30-billion-parameter model that generated tools on demand. Its tools also helped a 350-million-parameter model solve new problems. Reported inference-token use fell from 3,206 to about 100 when reusable tools replaced repeated step-by-step reasoning. These results strengthen the case for agent systems whose value lies in reusable execution logic and validation, rather than model size alone. They do not establish equivalent savings in enterprise deployments.
For builders, the concrete opportunity is to separate tool creation from routine execution, then measure the full cost of both. Appier describes potential marketing applications in which agents share tools and business rules across customer data, personalization and ad buying. The deployment hurdle is verifying generated tools when business tasks lack automatically checkable answers. Investors should look for evidence that validation remains reliable outside the research tasks and that savings survive tool maintenance and changing workflows.