Orchestrator displays 'explanation_of_tool_call' and 'new_instruction' to end-user

Question

Problem Summary:

• The agent includes technical fields such as explanation_of_tool_call and new_instruction in its responses.
• This occurs despite explicit "Instructions" configured in the agent to never show JSON code or internal thought processes.
• When attempting to bypass this by setting After running to "Send specific response (specify below)", the agent sends duplicate answers: first the specified variable, immediately followed by a secondary message from the orchestrator.
• The issue appears to be related to how the orchestrator handles "Rationale" and decision logic when executing tools.

Questions:

Why is the orchestrator ignoring "negative instructions" regarding JSON and instead dumping internal fields like explanation_of_tool_call to the user?

How can I configure the agent to only return the content of a specific output variable (String) without the orchestrator adding its own generative follow-up or technical rationale?

Is this a known limitation of "Generative Orchestration," and is there a way to force the agent to strictly provide plain-text output only?

Answer 1

Hello Miller, Rokas,
The issues you’re encountering are due to Copilot Studio’s Generative Orchestrator, not your agent’s prompts or configuration. The JSON fields and duplicate responses are system-generated planning metadata produced outside the LLM, so your instructions (including “do not show JSON”) cannot suppress them. When using “Send specific response”, the orchestrator still appends its own natural-language summary, resulting in duplicate messages—a known, non-configurable limitation where orchestrator output can only be appended, not replaced. The only reliable ways to return a single clean response are to end the topic with a return value, bypass orchestration via Power Automate, disable generative reasoning for deterministic flows, or strip metadata in a backend step; prompt engineering and negative instructions will not work.

Q1. Why is the orchestrator ignoring my instructions?

Because the leaking fields were not produced by the LLM — they are generated by the Copilot Orchestrator, which your instructions do not control.

Q2. How do I force the agent to output ONLY my variable?

Use one of these:

End the topic with “Return value” (best)

Run a Power Automate flow and return its output

Turn off generative orchestration for the topic

Strip orchestrator metadata using a custom API

Q3. Is this a known limitation?

Yes. It is a documented internal limitation of Generative Orchestration — tool‑planning metadata may be surfaced when:

Tool calls are executed

Instructions conflict with orchestrator defaults

Output is forced into deterministic mode