Looping

[ekjuanrejon]

I have created an agent that should ask the iser six questions and then create an xml. It ask question 1 then goes on question 2. Question 2 is answer and goes back to question 1. It is on a loop. Any ideas?

[moonbox3]

Can you provide more context, please? Even a code sample? It's hard for us to understand what could be going on just based on the original question.

[ekjuanrejon]

I cannot share because it is confidential but the gist is ask 6 questions and then create an xml.

Initially the agent was just creating the xml right away. Now it ask 2 questions and loops with questions 1 and 2.

Am thinking if there should be a triage agent

[ekjuanrejon]

Wondering if I should use a workflow and the six questions be the edges but how to store the answers until it arrives to the agent to create the xm

[ekjuanrejon]

You are a questionnaire assistant collecting 6 answers in strict order.

    PROCEDURE — follow these steps every turn:
    Step A: Count how many user answers already exist in the conversation history.
    Step B: Based on that count, ask ONLY the next question from this sequence:
      - 0 answers so far → ask for Full Name
      - 1 answer  so far → ask for Email Address
      - 2 answers so far → ask for Department
      - 3 answers so far → ask for Project Name
      - 4 answers so far → ask for Project Description
      - 5 answers so far → ask for Target Completion Date (YYYY-MM-DD)
      - 6 answers so far → call the GenerateXmlFile tool with all 6 answers
    Step C: Ask exactly one short question. Do NOT list all questions.
            Do NOT restart from the beginning. NEVER ask for a field you already have.

    HARD RULES:
    - Each user message after the first is an answer to the previous question.
    - Progress is strictly forward. Never go backwards.
    - Once you have 6 answers, call GenerateXmlFile immediately.
    
    something like that but the agent looks question at question 2.. keeps asking question 1 goes to question 2 and so

[ekjuanrejon]

testing that same prompt in foundry it works but does not work with MAF

[LeftTwixWand]

@ekjuanrejon which model do you use with MAF?

[ekjuanrejon]

Gpt 5 mini but was using dev ui to test. Not sure if that is the problem...

[msatfi89]

The problem is likely that the prompt relies on the model to verify its own history state ("Step A: Count how many user answers..."). LLMs are notoriously bad at counting logic within the prompt, especially if the framework (MAF) truncates or formats history differently than the Dev UI.

What fixed this for me was moving progress tracking out of the prompt and into a tool. Instead of asking the LLM to count, give it a store_answer tool that tracks state in Python and tells the agent exactly what to do next:

collected = {}

FIELDS = ["Full Name", "Email Address", "Department", "Project Name", "Project Description", "Target Completion Date"]

def store_answer(field: str, value: str):
    collected[field] = value
    remaining = [f for f in FIELDS if f not in collected]
    
    if not remaining:
        return "All 6 collected. Call GenerateXmlFile now."
    
    return f"Stored '{value}'. NEXT STEP: Ask the user for {remaining[0]}"

State should live in code, not in the prompt. That's why it works in Foundry (which likely feeds full context differently) but breaks here.

The other thing that helped me was adding a runtime safety net. I kept hitting these "Q1 → Q2 → Q1" loops across different models, so I built a small middleware called Aura Guard. It detects when the agent repeats the same sequence of tool calls/questions and breaks the cycle automatically. It's framework-agnostic (works with MAF), zero-dependency, and has a "shadow mode" if you just want to see what loops it would catch without interfering.