A worthy challenge? Have smolagents run the Proust questionnaire

[sahuguet]

The beauty of LLMs is the ability to have conversations.
I am trying to use smolagents to have a conversation with a human and extract some information, e.g. name, hobbies, etc. The agent should keep going until all the information has been gathered. The agent may choose to introduce questions or topics that are not relevant to the task, to make the conversation more interesting, etc.

A nice (and very French) way of framing this problem is the Proust Questionnaire.

When I try a simple prompt, smolagents generates a code snippets that terminates right away.

The code snippet should either implement a sequence of questions/answers; or should do a while True to provide some back and forth until the questions have been asked and the answers collected.

─────────────────────────────────────────────────── New run ────────────────────────────────────────────────────╮
│                                                                                                                │
│ You are **ProustBot**, a friendly conversational assistant inspired by the Proust Questionnaire.               │
│                                                                                                                │
│ Your job is to ask the user **exactly** these five questions, in this order, one at a time:                    │
│                                                                                                                │
│ What is your idea of perfect happiness?,                                                                       │
│ What is your greatest fear?,                                                                                   │
│ Which living person do you most admire?,                                                                       │
│ What is your greatest extravagance?,                                                                           │
│ Which talent would you most like to have?,                                                                     │
│                                                                                                                │
│                                                                                                                │
│                                                                                                                │
│ Guidelines:                                                                                                    │
│ — Begin with a brief, casual greeting and mention you'd like to ask five quick questions.                      │
│ — After each user response, acknowledge it very briefly (≤ 10 words) before asking the next question.          │
│ — Keep the tone warm, informal, and non‑judgmental.                                                            │
│ — After the fifth answer, thank the user and close the conversation gracefully.                                │
│                                                                                                                │
╰─ OpenAIServerModel - gpt-4.1 ──────────────────────────────────────────────────────────────────────────────────╯
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Step 1 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 ─ Executing parsed code: ───────────────────────────────────────────────────────────────────────────────────────
  first_message = (
      "Hi there! 😊 I’d love to ask you five quick questions—just for fun. "
      "Here’s the first:\n\n"
      "What is your idea of perfect happiness?"
  )
  final_answer(first_message)

Out - Final answer: Hi there! 😊 I’d love to ask you five quick questions—just for fun. Here’s the first:

What is your idea of perfect happiness?
[Step 1: Duration 2.60 seconds| Input tokens: 2,140 | Output tokens: 98]

[sahuguet]

Here is a solution using a white True loop. But I am curious if this is something we could more elegantly using smolagents.
See https://github.com/sahuguet/agentic-stuff/blob/main/agent_loop_01.py .