The Agent Class: Tracking Conversation History

Before implementing any ReAct-specific logic, this course's from-scratch walkthrough starts with something more basic: a minimal Agent class that wraps around a conversational LLM and keeps track of its full message history. This class doesn't know anything about Thought/Action/Observation yet — it's just the conversational memory foundation everything else gets built on top of.

The class holds two important pieces of state. system, an optional system prompt string that sets the agent's personality and behavioral constraints — if provided, it becomes the very first message in the conversation, using the special 'system' role, exactly as in the OpenAI Chat API format. And self.messages, a list that acts as the agent's conversation memory: every interaction, whether it's a user message or the assistant's own output, gets appended to this list. This history matters because LLMs are stateless between API calls — without re-sending the full message history on every turn, the model has no way to behave coherently across multiple turns.

The class's __call__ method is the core interface for interacting with the agent, and it does three things in one call: it records the incoming message by appending it to self.messages as a 'user' message, it calls a separate invoke() method that sends the entire conversation history to the LLM and gets back a reply, and it appends that reply to self.messages as an 'assistant' message before returning it to the caller. invoke() itself just handles the actual API call — sending the full message list and extracting the text content from the response.

This tiny class is deliberately unopinionated about ReAct — it's a general-purpose conversational memory wrapper. The ReAct behavior comes entirely from the system prompt you pass into it, which is exactly what the next subtopic covers.

from litellm import completion

class Agent:
    def __init__(self, system: str = ""):
        self.system = system
        self.messages: list[dict] = []
        if self.system:
            self.messages.append({"role": "system", "content": self.system})

    def __call__(self, message: str = "") -> str:
        if message:
            self.messages.append({"role": "user", "content": message})
        result = self.invoke()
        self.messages.append({"role": "assistant", "content": result})
        return result

    def invoke(self) -> str:
        response = completion(model="openai/gpt-4o", messages=self.messages)
        return response.choices[0].message.content

# Quick sanity check — this alone already "remembers" prior turns:
agent = Agent(system="You are a helpful assistant.")
print(agent("My name is Alex."))
print(agent("What's my name?"))  # correctly recalls "Alex" from self.messages