lagent

Why we included this project

Lagent borrows the layer-and-message-passing ideas from PyTorch and applies them to agent design, so composing a model, a set of tools, and memory works much like assembling a network in Python. You wrap a vLLM-hosted model in an Agent and let typed AgentMessage objects flow between components, while memory is taken care of on each call. It also ships practical pieces like tool parsers and aggregators for few-shot formatting, plus multi-agent message passing, so you can get a working agent up quickly and swap in your own behavior later. InternLM models are the natural starting point, but the same interface drives Qwen and other checkpoints.