Agents 2.0 Documentation

Agents 2.0 introduces a framework for training language agents through symbolic learning, inspired by connectionist learning (neural networks).

Key concepts include:

• Agent Pipeline as Computational Graph: The agent pipeline corresponds to a neural net's computational graph.
• Nodes as Layers: Each node in the pipeline acts as a layer.
• Prompts and Tools as Weights: The symbolic components (prompts and tools) within a node function like weights.
• Language-based Learning: The framework implements a "forward pass" (agent execution) to create a trajectory, calculates a "language loss" via prompt-based evaluation, performs "back-propagation" to generate "language gradients" (textual reflections), and updates symbolic components and the computational graph using specialized prompts.

You can install the Agents library directly from the GitHub repository using pip. Choose the method based on whether you want to use it as a dependency or develop it locally.

# Installation from git repo branch
pip install git+https://github.com/aiwaves-cn/agents@master

# Installation for local development
git clone -b master https://github.com/aiwaves-cn/agents
cd agents
pip install -e .