OctoTools
Open-source, training-free agentic framework developed by Stanford University researchers for complex reasoning across diverse domains:
• Leverages a “planner + executor” architecture with standardized tool cards to orchestrate high-level planning and low-level execution
• Tackles diverse benchmarks such as MathVista, MedQA, MMLU-Pro, and GAIA-Text with strong accuracy gains over baseline agents
• Integrates tools easily without retraining—tool cards define metadata for web search, code execution, image processing, and more
• Modular design includes vLLM support, interactive Playground demos, and PyPI package for streamlined development
• Winner of “Best Paper” at NAACL 2025 and designed for extensibility and reproducibility in research and production