Share
Report
2I've been working in n8n for a while creating myself an agent, simple loop, gpt4.1 with an elaborate system prompt and some mcp tools connected.
Mcp tools provide agent with access to crm/wms to ask about orders/products/status/tracking
IT WORKS FANTASTIC. I thought I've got staff some copilot licences, let's try and deploy it behind auth as a copilot agent so staff can use it.
I've tried to recreate the same thing on copilot studio. Same model, same system prompt, same mcp tools. Generative AI enabled.
It's dumber. It's missing context that was provided by tool responses. It's coming back with very dry and short responses. It delivers bare minimum in response. It always adds "If you need further assistance, feel free to reach out to customer support on the respective platform." with each response. It's not eager to call multiple tools to get the final answer.
Basically this, when asked for can you tell me what's the situation with this order [order number]:
n8n agent
goes and get's order content, matching pos for etas, looks up suppliers, suggest alternatives to non in stock items . tracking numbers and writes a complete nicely formatted here is the situation response. Suggest potential next steps to take to resolve any issues.
Copilot agent
The order is awaiting shipping. This is the items that are on this order LIST OF PRODUCT CODES, and that's it, not even qty - code - description(which is all provided), just codes.
It's like temperature is all wrong and there seems to be additional layer between what the agent generates and what the final response is going to be.
I've adjusted prompt, moderation level, response formatting, tried modifying the topics to see if it changes it. Nothing works.
And it's not like n8n is doing something special to make it happen, I've recreated the n8n behaviour in langchain, system prompt + mcp tools, literally 30 lines. Works as expected.
All responses (
Answers (
