Power Platform Community Forum Thread Details

Hi All - I am reaching out to discuss an issue I’m expecting with the primary Agent orchestration. I have around 8-10 sub agents with defined instructions and no knowledge base to extract the necessary information based on their instructions and I have linked all these sub agent to my primary agent - None of the sub agents have knowledge source and solely rely on the instructions and model selection.

in the Primary agent I’ve added the untied knowledge base ( root folder -share point) - in Sharepoint, for each company I’ve created metadata tags for both the folder and files within that folder

The Primary Orchestration should work like this: when an Analyst or user the enter the prompt with the client name as a keyword, the primary agent should recognise the keyword(company), search the knowledge source on share point, and then invoke each sub agent to extract the output as per the instructions. This process should be applied to all sub-agents, and primary agents should provide the raw output generated by each sub agent.

Especially, the primary agent should act as the orchestration agent , delivering the exact raw output generated by each sub agent. If anyone has insights or suggestions on how to resolve this issue, your help would be greatly appreciated.

currently using model - Claude Opus 4.5

Thank you for your attention and looking forward to your feedback.

Your orchestration scenario is a strong use case for Copilot Studio’s agent orchestration, but there are some important architectural and behavioral considerations:

Key Points & Recommendations
1. Primary Agent Knowledge Source vs. Sub-Agents
Primary agent: Has a SharePoint knowledge base (with metadata).
Sub-agents: No knowledge base, only instructions and model selection.
Behavior:

When a user prompt includes a company name, the primary agent will search the SharePoint knowledge base for relevant content.
Sub-agents, when invoked, will only use their instructions and the model—they will NOT access the knowledge base unless explicitly configured.

Implication:

Sub-agents will not “see” or use the SharePoint search results unless you explicitly pass that content to them as part of the orchestration logic.

2. How to Orchestrate Correctly
Recommended Pattern:

Primary agent receives the user prompt.
Primary agent searches SharePoint for content matching the company keyword (using metadata).
Primary agent extracts or summarizes the relevant content.
Primary agent passes the relevant content (or a reference to it) as context/input to each sub-agent, along with the sub-agent’s specific instructions.
Each sub-agent processes the provided content according to its instructions and returns its output.
Primary agent aggregates and returns the raw outputs from all sub-agents.

Why?

Sub-agents without a knowledge base will not perform retrieval-augmented generation (RAG) or search SharePoint on their own. They need the primary agent to “feed” them the relevant content.

3. How to Implement in Copilot Studio

Use orchestration flows: In the primary agent, after retrieving content from SharePoint, use orchestration logic to call each sub-agent and pass the retrieved content as an input variable.
Prompt engineering: For each sub-agent, construct a prompt like:
“Given the following company information: [SharePoint content], perform [sub-agent instruction].”
Aggregate results: Collect the outputs from each sub-agent and return them as a combined/raw output to the user.

4. Model Selection

Claude Opus 4.5 is capable, but ensure your orchestration logic is clear and that each sub-agent’s instructions are precise.
If you want sub-agents to use knowledge sources in the future, you must assign them explicitly.

5. Potential Pitfalls

If you do not pass the SharePoint content to sub-agents, they will only use their instructions and the user’s original prompt, likely missing key context.
Metadata tagging in SharePoint is helpful, but ensure your primary agent’s search logic leverages these tags for accurate retrieval.