agents

Why we included this project

Most agent frameworks leave you hand-tuning prompts until the behavior sticks. This one borrows an idea from how neural networks are trained, treating the agent pipeline like a computational graph with prompts and tools as its weights. You run the agent on a task, the framework scores the outcome, then sends textual reflections back through each step to update the prompts and tools, much the way back-propagation updates weights. That feedback loop lets teams keep improving agent behavior on their own data rather than starting over each time, and it extends to multi-agent setups where several agents share a pipeline. The 2.0 release added evaluation tooling as well, so this suits researchers and engineers prototyping agents, not people who want a quick chatbot wrapper.