Agents is an open-source framework from AIWaves for building data-centric, self-evolving autonomous language agents. Its distinctive idea is symbolic learning: treating an agent pipeline like a neural network so the agent itself can be trained and improved, not just prompted. The project has about 6,000 stars on GitHub and is released under the Apache 2.0 license.
The framework draws a deliberate analogy with connectionist learning. An agent pipeline corresponds to a computational graph, each node in the pipeline to a layer, and the prompts and tools at a node to the layer’s weights. With that mapping, the machinery of neural training — forward evaluation, language-based loss, textual “gradients” and weight updates — can be applied to optimize the agent. Version 2.0, released with an accompanying research paper, added this agent learning loop plus evaluation support, moving the project beyond a static agent-building library toward agents that improve from data and feedback.
The project comes from the AIWaves research group and is documented on ReadTheDocs, with a project page and an associated arXiv paper describing the symbolic learning framework. Development has been quieter since the 2.0 release in 2024 — the last repository push was in September 2024 — so it is best treated as a research framework to learn from and build on rather than a fast-moving platform. Its core ideas, especially optimizing agents with language-based feedback loops, have nonetheless become a recurring theme across agent research since.
Free. Agents is open source under the Apache 2.0 license; you pay only for the LLM API calls your agents make.
Researchers and developers experimenting with trainable, self-improving LLM agents, especially those interested in systematic optimization of prompts and tools.