Environment ​

Early MetaGPT versions used the same implementation of Environment in metagpt/environment/base_env.py to describe the environment of a software team. In order to support the application of multiple agents in different environments (such as games, mobile phones, etc.), we have expanded the overall definition and functions of the environment to better support the development of different scenarios.

Definition ​

ExtEnv external environment is defined as the carrier for the interaction between the agent and different environments. For example, the game to be connected by the agent is a remote engine service and a collection of API interfaces provided externally. We manage the docking API by defining the external environment, so that the agent has the direct ability to interact with the game. For another example, the agent connects to the role-playing Werewolf game and defines the external environment to implement the Werewolf game rules itself, so that the agent can know the current status of the character after dark and dawn.

The internal environment of XxxEnv is defined as the environment used directly by the agent and its team, similar to the Environment used in software development. Generally, it directly inherits ExtEnv and adds additional custom parts as required.

In ExtEnv, we refer to the design of gymnasium in the reinforcement learning scenario and introduce the observation_space and the action_space, which are used to represent the set of states obtained from the external environment and the action collection that can act on the environment respectively. Through the definition of two spaces, you can simplify the abstraction in different environments, allowing the environment provider to focus on the implementation of environment logic and the user to focus on the use of different space values.

In addition, the decorators mark_as_readable and mark_as_writeable for the different read-write interfaces provided by ExtEnv are also provided to facilitate the unified management of method interfaces for external environment docking, so that subsequent agents can use them as A tool capability that can directly and automatically call different external environment docking interfaces based on the input natural language (this part of the function is to be opened).

Different Environments ​

Currently, we provide several scenario environments and provide corresponding scenario usage entrances under MetaGPT/examples/.

• ToBeAdded, Minecraft Environment
• Added, Werewolf Environment
• Added, Stanford Town Environment
• Added, Android Environment
• ToBeAdded, Web Environment