This example demonstrates how to use AgentScope to play the Werewolf game, where six agents play against each other as werewolves and villagers. You will learn the following features in AgentScope:
- How to make an agent respond with different specified fields in different situations in AgentScope (Like a finite state machine!)
- How to use msghub and pipeline in AgentScope to construct a game with COMPLEX SOP
The werewolf game involves a complex SOP with multiple roles and different phases. In these phases, players with different roles should take different actions, e.g. discussion, voting, checking roles (seer), using potions (witch), and so on.
Therefore, for an agent in werewolf game, it should be able to switch its status according to the game phase and its role, and respond accordingly. Of course, we can hard code the SOP (or finite state machine) within the agent, but we expect the agent to be more flexible, intelligent, and adaptive, which means the agent shouldn't be designed for a specific game only. It should be able to adapt to different tasks and SOPs!
To achieve this goal, in AgentScope, we use a built-in DictDialogAgent class, together with a parser module to construct the werewolf game.
Moreover, the pipeline and msghub in AgentScope enable us to easily construct a complex SOP with multiple agents. We hope the implementation is concise, clear and readable!
More details please refer to our tutorial:
Werewolf-running-example.mp4
This example has been tested with the following models:
- dashscope_chat (qwen-turbo)
- ollama_chat (llama3_8b)
To run this example, you need to:
- Set your API key in
./configs/model_configs.json
[Optional] To play the game in person, see werewolf.py for the complete code.
The following code is the implementation of daytime discussion in Werewolf. With msghub and pipeline, it's very easy to program a discussion among agents.
# ...
with msghub(survivors, announcement=hints) as hub:
# discuss
x = sequentialpipeline(survivors)
# ...The parser is used to specify the required fields in the agent's response and how to handle them. Here's an example of a parser configuration:
to_wolves_vote = "Which player do you vote to kill?"
wolves_vote_parser = MarkdownJsonDictParser(
content_hint={
"thought": "what you thought",
"speak": "player_name",
},
required_keys=["thought", "speak"],
keys_to_memory="speak",
keys_to_content="speak",
)In this example, the MarkdownJsonDictParser is used to parse the agent's response. The content_hint parameter specifies the expected fields and their descriptions. The required_keys parameter indicates the mandatory fields in the response. The keys_to_memory and keys_to_content parameters determine which fields should be stored in memory and included in the content of the returned message, respectively.