CaptainAgent requires nested_config for configuration. Below is an example, we will break it down and provide a detailed explanation.

nested_config = {
    "autobuild_init_config": {
        "config_file_or_env": "OAI_CONFIG_LIST",
        "builder_model": "gpt-4o",
        "agent_model": "gpt-4o",
    },
    "autobuild_build_config": {
        "default_llm_config": {"temperature": 1, "top_p": 0.95, "max_tokens": 1500, "seed": 52},
        "code_execution_config": {"timeout": 300, "work_dir": "groupchat", "last_n_messages": 1},
        "coding": True,
        "library_path_or_json": "captainagent_expert_library.json",
    },
    "autobuild_tool_config": {
        "tool_root": "default",  # this will use the tool library we provide
        "retriever": "all-mpnet-base-v2",
    },
    "group_chat_config": {"max_round": 15},
    "group_chat_llm_config": llm_config.copy(),
}

If you run CaptainAgent without specifying a nested_config, a default is used and this requires an OAI_CONFIG_LIST.

The default model used is OpenAI’s GPT-4o.

To see the default configuration and the OpenAI model used, see the captainagent.py file.

For advice on using OAI_CONFIG_LIST, see the documentation.

​
autobuild_init_config

This section is used to configure the initial setup of autobuild. autobuild_init_config takes in arguments from AgentBuilder.__init__(). Check the full list of arguments here.

​
config_file_or_env

Configures the path to API key config. Defaults to OAI_CONFIG_LIST.

​
builder_model

Configures the backbone of agent builder. The builder is used for agent selection from the library. Defaults to gpt-4o.

​
agent_model

Configures the backbone of agents in the group chat. Defaults to gpt-4o.

​
Other kwargs

autobuild_init_config takes in arguments from AgentBuilder.__init__(). Check the full list of arguments here.

​
autobuild_build_config

This section is used to configure the building process of autobuild. autobuild_build_config takes in arguments for AgentBuilder.build(). Check the full list of arguments here.

​
default_llm_config

Configures the default parameters for the builder during the autobuild process. Defaults to {"temperature": 1, "top_p": 0.95, "max_tokens": 2048}. config_list is not needed here.

​
code_execution_config

Configures how the user proxy executes code within the nested chat. Defaults to {"timeout": 300, "work_dir": "groupchat", "last_n_messages": 1, "use_docker": False}. Full configuration docs here.

​
coding

Enables or disables whether to add the user proxy in the nested chat. Defaults to True.

​
library_path_or_json

Specifies the path to the agent library file. For details on customizing your own agent library, refer to the agent library page.

​
kwargs

autobuild_build_config takes in arguments for AgentBuilder.build(). Check the full list of arguments here.

​
autobuild_tool_config

This section is used to configure how to retrieve the tool library for the agents in the group chat. For details on customizing your own tool library, refer to the tool library page.

​
tool_root

Specifies the root directory of the tool library. When set to 'default', it will load the default library we provide.

​
retriever

Configures the retriever model used for fetching relevant tools from the library. Defaults to all-mpnet-base-v2. The value is valid as long as it is found by SentenceTransformers library.

​
group_chat_config

This section is used to configure the group chat settings. group_chat_config also takes in arguments for initializing autogen.GroupChat. Refer to all the configurables here.

​
max_round

Specifies the maximum number of rounds in a group chat session. Defaults to 10.

​
kwargs

group_chat_config also takes in arguments for initializing autogen.GroupChat. Refer to all the configurables here.

​
group_chat_llm_config

Specifies the LLM config of the GroupChatManager.

Agent LibraryTool Library