DynaSaur is a dynamic LLM-based agent framework designed for complex reasoning and task execution, particularly in scenarios requiring adaptation beyond predefined actions. It targets researchers and developers building sophisticated AI agents capable of self-improvement and handling novel situations, offering a significant advantage in benchmarks like GAIA.

How It Works

DynaSaur leverages a programming language, specifically Python snippets, as a universal action representation. At each step, the agent generates Python code that can either invoke existing actions or dynamically create new ones by composing or developing them from scratch. This approach allows the agent to expand its capabilities organically, improving its ability to recover from failures or address situations where no pre-existing actions are suitable.

Quick Start & Requirements

• Install: Clone the repository and install dependencies via pip install -r requirements.txt after setting up a Python 3.12 Conda environment.
• Prerequisites: Requires API keys for Azure OpenAI (for main agent and embeddings) and optionally for SerpApi and Azure GPT-4V. GAIA benchmark dataset download is also necessary, requiring Hugging Face CLI login.
• Setup: Environment setup and data download are estimated to take 15-30 minutes depending on internet speed and data size.
• Links: GAIA Benchmark

Highlighted Details

• Achieves top performance on the GAIA benchmark, leading non-ensemble methods.
• Dynamically generates Python code to call or create new actions.
• Demonstrates versatility in recovering from action failures or missing capabilities.
• Supports gradual expansion of a reusable action library.

Maintenance & Community

The project is from Adobe Research. A TODO item indicates future support for the OpenAI API. Citation details for the associated paper are provided.

Licensing & Compatibility

The repository does not explicitly state a license in the README. Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

The project is marked with TODOs, indicating ongoing development, specifically the planned addition of OpenAI API support. The absence of an explicit license may pose restrictions for commercial adoption.