Power Platform Community Forum Thread Details

Question about the new knowledge management in the preview of the multi-agent

hello everybody,

A question for those who play in preview environnement (public one, enabled at env creation).

Since a moment i noticed something strange and new : the knowledge is no more "search" inside "sub agent" / "child agent" but directly from "master" agent : here is a screen to illustrate :

here is my agent organisation everytiime i ask something

how is it organise :

Everybody with GPT 5 Auto

Master agent :

- no knowledge

- no tool

- no instruction

- agents

- overview : General knowledge disable (if enable it loose reference source from child and loose ability to use sub agent, use directly sources); web search disable

-- 3 sub agent

In each sub agent

- around 20 SharePoint knowledge source with their own description and around 25 topic with their own prompt

The "2 search sources" are search inside both sub agent, but "after" and out of context of the sub agent (so not using any rulez or parameters from the sub agent)

if i make the test in "normal" environnement (not preview) the knowledge is inside the sub agent

Where is the problem :

It produce totally different answer :

Sub agent 1 : Ak question 1 : answer ok

Ask the question from master agent -> answer totally different and wrong and use different source (we made dozen of test)

How can you help me ?

Are you aware of new behavior inside the multi agent in preview ? some news about it to understand ? or a m i not understanding the general knowledge, or "pass full contexte or not" to the agent ? any tips woulb be appreciated :)

EDIT : i forgot : but i have topics inside sub agent and i have "question" wich trigger prompt from the sub agent, but the same question from "master agent" never trigger the prompt inside the child agent...

Categories:

General

Hi @ Romain The Low-Code... ,

What you’re seeing in the preview environment is how generative orchestration + multi-agent currently works: the master agent’s planner can call knowledge sources directly (shown as “Search sources”) instead of always handing off to the child agent and letting the child do the knowledge search.

Microsoft describes this behavior in the generative-orchestration docs:

With generative orchestration, an agent can choose “the best tools, knowledge, topics, and other agents to answer user queries.” Child/connected agents and knowledge sources are all just options the planner can pick from.
When generative orchestration is enabled, knowledge search is done by the planner and doesn’t go through the classic Conversational boosting topic or any customizations you made there.

That’s why in your activity map you see:

The master agent,
Then “Search sources” calls,
And no steps taken inside the sub-agent, even though the same question works correctly when you talk to the sub-agent directly.

In other words: in the preview, the master agent is allowed to “skip” the sub-agent’s own routing logic and go straight to the underlying knowledge.
How to get more predictable behavior

If you want strict routing through the sub-agent (classic behavior)
Switch the master (and/or child) agent to classic orchestration. In classic mode, topics and redirects control the flow instead of the generative planner:
- Go to the agent’s Settings → Generative AI → Orchestration and set Use generative AI orchestration to No.
- Use an explicit Redirect to agent node in a topic so that the master hands off to the child, and the child’s own topics + knowledge run in context.
If you want to stay in generative orchestration (preview behavior)
Then the planner will continue to choose among: topics, tools, other agents, and all available knowledge sources. To improve routing:
- Give each child agent and each knowledge source a very clear description that states when it should be used and what domain it covers. The planner uses these descriptions as the main signal when picking what to call.
- Optionally reference the child agent in the master’s Instructions (using /childAgentName) so the planner knows that specific types of questions should be handled by that agent.
Remember this is preview behavior
Microsoft explicitly notes that the “Add other agents” article and multi-agent features are preview and subject to change, so behavior in the preview environment may differ from non-preview environments and can evolve over time.