Agent ignores instruction
I’m trying to build a Copilot Studio agent to help create RFQs in our ERP system. The ERP is connected through MCP and exposed as a tool in the agent.
The agent also has a fairly large and structured instruction set (~7k characters) describing:
how the agent should behave
when different MCP tools should be invoked
how the workflow should work step-by-step
There is no Knowledge Base configured, since all required logic and instructions are already included in the Instructions section. I’ve also disabled both web search and general answers. The model is GPT-5 Chat.
The problem is that the agent behaves very inconsistently. Occasionally it works exactly as expected — it creates a reviewable suggestion and then successfully creates the RFQ in the ERP system. However, around 90% of the time it seems to completely ignore the instructions. After ~20 seconds it responds with something like:
“I cannot help with this task”
or
“There is no knowledge available regarding this topic.”
This makes me wonder if I’m approaching this the wrong way.
A few questions:
Is this a common issue with Copilot Studio orchestration + MCP tools?
Are long instruction prompts generally unreliable in this scenario?
Would it be better to disable orchestration and instead build the flow using Topics?
Are there best practices for improving tool reliability and instruction-following in Copilot Studio agents?
Any suggestions or experiences would be greatly appreciated.