Inconsistent Chat Agent Behavior with Tool Execution, Field Collection, and Normalization

Questions
1

Describe the problem/error/question

I’m experiencing persistent and inconsistent behavior in an AI agent–driven chat workflow.
Despite very explicit prompt rules, the agent does not reliably follow required steps, enforce data normalization, or execute all required tools in sequence.

The workflow is designed for a maintenance intake use case and requires the agent to:

  • Ask for required information in a strict order

  • Normalize user input into specific data types (arrays, IDs)

  • Execute multiple tools in a required sequence (Create / Update Ticket → Notify Super)

In practice, the agent behaves inconsistently between runs, even with the same prompts and conversation structure.

What is the error message (if any)?

There is no single hard error message in most cases.
Instead, the issues manifest as silent logical failures, including:

  • Tools not being executed when required

  • Tools executing with malformed or missing inputs

  • Downstream workflow steps breaking due to incorrect data types or missing identifiers

Please share your workflow

https://n8n-latest-qvm6.onrender.com/workflow/uQaBucd3zdN7H7fN

Core Issues Observed

Core issues observed

  1. The agent reports updates that did not actually occur
    In some cases, the agent responds conversationally as if an existing ticket was successfully updated, but the update did not occur in the workflow at all.

  2. New tickets are created without required questions being asked
    The agent sometimes creates a new ticket without first asking for:

    • A detailed description of the issue

    • Preferred service days

    • Preferred service time windows
      This happens even when:

    • The information was never asked for earlier in the conversation, and

    • The prompt explicitly requires this information before ticket creation.

  3. Normalization rules are not consistently enforced before tool execution
    In some runs, the agent passes values such as:

    "Monday, Tuesday"

    instead of the required normalized format:

    ["Monday", "Tuesday"]

    This breaks downstream nodes that strictly expect arrays.

    In another case, the agent passed the unit_id into a follow-up tool in an incorrect format, which caused that tool to fail and disrupted the workflow.

  4. Adding guardrails causes other required steps to be skipped
    When stricter rules are added (for example, requiring scheduling information before ticket creation), other required steps are sometimes skipped instead — such as asking the tenant to describe the issue in more detail.

  5. Required follow-up tools are not reliably executed
    When an existing ticket is updated and scheduling information changes, the agent is supposed to call an additional follow-up tool (for example, “Notify Super”).
    In practice, the agent sometimes updates the ticket but does not execute the required follow-up tool, even though the prompt explicitly requires it.

Overall, this makes it feel like:

  • Tool execution order is not guaranteed

  • Required tools in a sequence can be silently skipped

  • Data passed between tools is not reliably preserved or validated

Information on your n8n setup

  • n8n version: 2.0.3 (Self-hosted)

https://gist.github.com/getconversationalai/8e522d5060330397ed910a0bae98d040

2

hey @development welcome to n8n community.

try changing ai model(llm)

3

Hi @development,

Please can you share your workflow so we can have a look at your system prompt. There are a few tricks to improve the system prompt to follow your rules

4

https://gist.github.com/getconversationalai/8e522d5060330397ed910a0bae98d040