Agent Class

Overview

The Agent is an abstract class for creating autonomous agents. These agents can perform actions and communicate with other providers. During instantiation, an agent is initialized with a set of instructions and can optionally be linked with a database provider.

An Agent works by advancing through a sequence of tasks. It implements iterator methods (__iter__ and __next__) for this purpose. Each iteration corresponds to a step of the task that the Agent has been designed to accomplish. This step could be a conversation turn, which involves generating a new message from the ‘assistant’ and then parsing the reply from the ‘user’. The run method can be used to execute these tasks until completion, with the task being deemed complete when the __next__ method returns None.

It has abstract properties for fetching its responses, associated conversation, and tools, whose concrete implementation is instantiated by subclasses. It also has an abstract method for setting a database provider, essential for managing conversations with the user.

Usage Example:

The following example shows a basic creation of a subclass of Agent:

class SimpleAgent(Agent):
    """Implements the abstract Agent class for a simple specific agent."""

    def __init__(self, instructions: str) -> None:
        super().__init__(instructions)

    def __iter__(self):
        ...

    def __next__(self) -> str:
        ...

    @property
    def conversation(self) -> LLMConversation:
        ...

    @property
    def agent_responses(self) -> List[LLMChatMessage]:
        ...

    @property
    def tools(self) -> Sequence[Tool]:
        ...

    def run(self) -> str:
        ...

This example shows a simple implementation of the Agent abstract class. The ... represents sections of code that must be implemented to define the specific behaviour of the SimpleAgent.

Related Symbols

• LLMChatMessage, LLMConversation: Models for handling and representing chat messages and conversations.

• Tool: An abstraction for different types of tools associated with the agent.

• LLMConversationDatabaseProvider: Abstract base class for database providers.

Limitations

The Agent abstract class doesn’t provide an easy method to modify or control the flow of execution. It assumes that all tasks are to be performed in a cyclical manner and that they complete after a specific number of steps.

Follow-up Questions:

• How to handle more complex workflows that require non-linear execution paths?

• Is it possible to dynamically adjust the maximum number of iterations based on the task complexity?