OctoTools
Summary: OctoTools (Lu et al., Stanford, 2025) is a training-free open-source agentic framework for complex reasoning. Three components — tool cards, planner, executor — enable LLMs to use external tools across diverse domains without retraining. Achieves 9.3% average accuracy gain over GPT-4o across 16 benchmarks.
Sources: Academia/_mS6uP.pdf
Last updated: 2026-05-06
Citation
Lu, P., Chen, B., Liu, S., Thapa, R., Boen, J., & Zou, J. (2025). OctoTools: An agentic framework with extensible tools for complex reasoning. Stanford University. arXiv:2502.11271.
Core architecture
- Tool cards: standardized metadata encapsulating each tool’s functionality, inputs, and outputs. Enable plug-and-play integration without additional training or framework modification.
- Planner: two levels — high-level planning (global objective decomposition) and low-level planning (step-by-step action refinement). Includes a Context Verifier that decides whether to continue or terminate.
- Executor: instantiates tool calls as executable commands; stores structured results in context for the planner.
Performance
Tested on 16 diverse benchmarks including MathVista (visual math), MMLU-Pro (multi-domain), MedQA (medical), and GAIA-Text. OctoTools outperforms AutoGen, GPT-Functions, and LangChain by up to 10.6% when given the same tools.