Open source platform OctoTools has been released by scientists at Stanford University, who claim it can improve the performance of large language models (LLMs) for reasoning tasks by breaking them down into subunits.
The agentic platform enhances models with tools and can be extended with bespoke tools and workflows, and outperformed other LLM application frameworks, according to its developers.
OctoTools uses a modular approach to planning and reasoning, including tool cards which act as wrappers for tools including Python code interpreters and web-search APIs, as well as an optimisation algorithm to select the best tools for each task.
Its planner module generates a high-level plan based on the required skills and tools identified by feeding in a new prompt, which is then refined by an action predictor module.
A command generator maps the text-based plan to Python code to invoke the relevant tools, which are then run by the command executor.