NeuralStackAppier’s research paper on AI agents that build and use their own tools has been accepted at NeurIPS, a leading conference in artificial intelligence and machine learning. The paper introduces SMITH, short for Schema-grounded Multi-task Iterative Tool Honing, a reinforcement-learning framework designed to train both capabilities in a single loop.
The central problem is that AI systems may be able to create tools without reliably using them. In many existing workflows, engineers build APIs or configure fixed tools in advance, while tool creation and tool use are handled separately. Appier says this separation makes it difficult for the system that creates a tool to learn whether its description is clear, its parameters are well designed, or the tool works reliably in practice.
SMITH links those stages through a cycle of building, using, verifying and refining. When an agent encounters a vague description, incorrect parameters or a failed execution, the result feeds back into training. The framework therefore treats successful use on real problem-solving tasks as part of the process for improving the tools themselves.
Appier’s study uses a staged setup that begins with four simple examples. The system then evaluates its tools on 16 more difficult problems it has not previously seen, retaining tools that solve new tasks and adding them to a shared library. Higher-quality tools replace weaker ones, allowing multiple AI agents to reuse skills rather than rebuild them for every assignment.
The company reports that a model with about 4 billion parameters trained with SMITH created tools that outperformed those produced by every other method in the study on unseen tasks. It also surpassed a comparison in which a model with roughly 30 billion parameters created tools on the fly. The finding suggests that effective tool creation may not depend solely on using the largest available model.
Those tools were also usable across model sizes, according to Appier. A lightweight model with about 350 million parameters handled new tasks with tools created by the smaller tool-building model, while larger models also benefited from the shared tools. Appier says this could let agents divide work across systems with different capabilities.
Why this matters
SMITH points toward an agent architecture in which problem-solving experience becomes reusable infrastructure. Instead of repeatedly spending tokens on the same chain of reasoning, an agent can turn a validated process into a callable tool and make it available to other models. In Appier’s experiments, average output fell from 3,206 tokens with conventional step-by-step reasoning to about 100 tokens, a roughly 32-fold reduction, while task performance was maintained.
That approach could make enterprise agent systems more scalable for repetitive operations such as converting financial metrics, processing data and querying reports. It also offers a path toward multi-agent collaboration in which smaller and larger models share proven capabilities rather than operating as isolated systems.
Appier says its researchers next want to develop models that can continuously interact with their environments and take on a wider range of tasks. The NeurIPS acceptance gives the company’s research a prominent venue as it works to apply the framework to real-world Agentic AI deployments.