This framework provides a systematic approach to training and evolving autonomous language agents, inspired by neural network learning procedures. It targets researchers and developers building complex agent systems, enabling self-improvement through symbolic learning, akin to gradient-based optimization in deep learning.

How It Works

The core innovation is "symbolic learning," which treats agent pipelines as computational graphs and prompts/tools as weights. It implements a connectionist learning analogy: a "forward pass" executes the agent, storing its trajectory. A prompt-based loss function evaluates the outcome, generating "language gradients" via back-propagation through the trajectory. These gradients then update the agent's symbolic components and graph structure. This method supports multi-agent system optimization by treating agents as nodes or allowing collaborative actions within nodes.

Quick Start & Requirements

• Install via pip: pip install git+https://github.com/aiwaves-cn/agents@master
• Local development: git clone -b master https://github.com/aiwaves-cn/agents && cd agents && pip install -e .
• Prerequisites: Python, Git. Specific LLM API keys or local model setup may be required for agent execution.
• Documentation: Docs

Highlighted Details

• Introduces "symbolic learning" for self-evolving language agents.
• Analogous to connectionist learning (forward pass, back-propagation, weight updates).
• Supports optimization of multi-agent systems.
• Version v2.0.0 released with learning and evaluation support.

Maintenance & Community

• Last commit: [Date of last commit]
• PRs Welcome.
• Related papers: Agents 2.0, Agents

Licensing & Compatibility

• License: Apache 2.0.
• Permissive license suitable for commercial use and integration into closed-source projects.

Limitations & Caveats

The framework is described as a "systematic framework" and "major update," suggesting ongoing development. Specific performance benchmarks or detailed comparisons to existing agent frameworks are not provided in the README.