Python API
agents
Llm Agent

llm_agent.py

This Python module includes several classes and utilities for implementing agents within an environment using asynchronous execution (asyncio). The key class, LLMAgent, leverages a large language model (LLM) to generate actions based on observations and goals. It utilizes the sotopia library, presumably for multi-agent systems, and can produce script-like outputs if specified. Other classes include HumanAgent for human input and ScriptWritingAgent for generating scripted actions.

Classes

LLMAgent

This class represents an agent that interacts with an environment using a language model (LLM) to generate actions based on observations and goals.

Constructor: __init__

Initializes an instance with several attributes:

Arguments:

  • agent_name: str | None, default None. Used for agent identification.
  • uuid_str: str | None, default None. A unique identifier for the agent.
  • agent_profile: AgentProfile | None, default None. An optional profile for the agent.
  • model_name: str, default "gpt-3.5-turbo". Specifies the LLM model used.
  • script_like: bool, default False. Determines if the agent produces script-like outputs.

Usage:

agent = LLMAgent(agent_name="My Agent", model_name="gpt-3.5-turbo")

Attributes:

  • model_name: str | None. The name of the LLM model, defaulting to "gpt-3.5-turbo".
  • script_like: bool. Determines if the agent produces script-like output.

Methods

  • goal(): Returns the agent's goal as a string. If not set, it raises an exception.

    Returns: str

    Usage:

    llm = LLMAgent()
llm.goal = "Write a poem about nature"

  • act(): Generates an action based on the agent's observations and goals.

    Arguments:

    • _obs: Observation. The current observation from the environment.

    Returns: AgentAction

    Usage:

    agent = LLMAgent()
action = agent.act(observation)

ScriptWritingAgent

This class extends the LLMAgent to generate actions by composing scripts based on observations using a language model.

Constructor: __init__

Arguments:

  • agent_name: str | None, default None.
  • uuid_str: str | None, default None.
  • agent_profile: AgentProfile | None, default None.
  • model_name: str, default "gpt-3.5-turbo".
  • background: ScriptBackground, required for initialization.

Usage:

script = ScriptWritingAgent(agent_name="User", background=ScriptBackground())

Attributes:

  • model_name: str | None, default "gpt-3.5-turbo".
  • agent_names: list[str], stores agent names.
  • background: ScriptBackground, required for script generation.

HumanAgent

This class simulates a human agent interacting with the environment by prompting the user for input and acting based on the user's choices.

Constructor: __init__

Arguments:

  • agent_name: str | None, default None.
  • uuid_str: str | None, default None.
  • agent_profile: AgentProfile | None, default None.

Usage:

agent = HumanAgent(agent_name="User", uuid_str=None)

Attributes:

  • model_name: "human". Specifies that this agent represents a human user.

Methods

  • goal(): Returns the current goal, or prompts the user for input if no goal is set.

    Usage:

    human_agent = HumanAgent()
goal = human_agent.goal

  • act(): Prompts the user to select an action based on observations and returns an AgentAction.

    Usage:

    human_agent = HumanAgent()
human_agent.act(obs={"available_actions": ["move", "jump"]})

Agents

This class manages a collection of agents and allows for resetting or performing actions across multiple agents.

Methods

  • reset(): Resets all agents in the collection.

    Usage:

    agents = Agents({"agent1": Agent(), "agent2": Agent()})
agents.reset()

  • act(): Collects actions from all agents based on their respective observations.

    Arguments:

    • obs: dict[str, Observation]

    Returns: dict[str, AgentAction]

    Usage:

    agents = Agents()
actions = agents.act(obs_dict)

Functions

ainput()

Asynchronously retrieves user input while other tasks run concurrently.

Arguments:

  • prompt: str, default "".

Returns: str

Usage:

async def main():
    response = await ainput("Enter your name: ")
    task_name = response
Base Agent API DocsRedis Agent