- StandardNode as a universal node for prompting llm for steps; - Computer tool - docker container with the Linux desktop exposed as the tool. Screenshot observations, mouse, keyboard, and cli commands as actions; - The agent and the environment can now be defined completely in the Hydra configuration.
0.1.4
Easy way to create complex environments with ToolCollectionEnvironment
- Form Filler Agent described in [paper](https://arxiv.org/abs/2412.08445) now available in the examples - uv as the package manager - Agent optimization examples - One agent can use many LLMs now - Faster RL loops with batching, multi-GPU and multi-node support - Using VLLM with log probability and token ids during RL training