Start from the action space
LLM tool-use and multi-turn agent tasks have different needs from continuous robotic control. Identify observations, action space, external tools, and desired evaluation loop before choosing a framework.
For LLM environments
Verifiers and NeMo Gym provide infrastructure for building and evaluating environments around language models and agents. Check the current APIs, licensing, and execution constraints in their official repositories.
For browser agents
BrowserGym standardizes access to several web-agent tasks and benchmarks. It is especially relevant when the observation and action loop happens inside a browser.
For classic and embodied RL
Gymnasium offers a common environment interface. For multi-agent research, PettingZoo provides multi-agent interfaces. Robotics often needs an additional simulation engine and extensive safety testing.
How to decide
Build the smallest reproducible task first. Only add distributed rollouts, long-horizon simulation, or training infrastructure after confirming that the verifier works.
Try the blueprint generator to structure a task, action space, and verifier before writing your first environment.
Generate a blueprint →