Run agent only on message list instead of user prompt string?

[Philipp-Userlike]

Hi, first thanks for building this awesome library!
I work with an existing chat app and it represents a chat as a list of messages, and the goal is the get a next message.
The last message is not always sent by a user, there are situations where you have an assistant message last and still want to generate a new one.

So doing this would be nice:

result = agent.run_sync(
    [
        {"type": "user", "content": 'Where does "hello world" come from?'},
        {"type": "assistant", "content": "Let me look it up..."},
    ]
)

I know I can format the chat into the user prompt like this:

messages = [
    {"type": "user", "content": "Where does "hello world" come from?"},
    {"type": "assistant", "content": "Let me look it up..."},
]
chat_formatted = "\n".join(
    f"{msg['type']}: {msg['content']}" for msg in messages
)
prompt = """
Please generate an assistant answer in this chat:
{chat}
"""
result = agent.run_sync(prompt.format(chat=chat_formatted))

But I'm not fully sure what downside this would have. I could imagine:

• Generation quality being worse, at least for some models
• We use an obeservability and tracing tool. I will likely represent the input worse in the UI. Currently I can also add observed generations to a dataset, where a list representation is also used

Would it be possible to allow the agent to run on a list of messages? Or is it already possible in some way and I missed it?

[Wh1isper]

There is a message_history parameter

        message_history: list[_messages.ModelMessage] | None = None,

And you can use result.all_messages() to get all history message or build it manually: https://ai.pydantic.dev/message-history/#accessing-messages-from-results

[Philipp-Userlike]

I know, but there is a required user_prompt string as well in the agent run method.

So I don't know what to do, if I want to run an agent for a chat, where the assistant wrote the last message.