mcscatblog/posts/copilot-studio-child-connected-agents-inputs-outputs/

[utterances-bot]

Defining Inputs and Outputs in Child and Connected Agents in Copilot Studio | Microsoft Copilot Studio CAT Blog

Connected and child agents

http://localhost:4000/mcscatblog/posts/copilot-studio-child-connected-agents-inputs-outputs/

[inescavem-ai]

Hello! Thank you for the insights.
If I define an output property of a Connected Agent - How do I reference it in formulas and nodes of the Parent Agent?

Thank you.

[johnsonmattMS]

There is probably a reason this isn't easy to do in the UI. Dewain Robinson is very passionate on this topic.
If you use a connected agent there is nothing to stop the owner of that agent changing what it does, changing it's knowledge sources or just turning it off altogether. This could introduce instability or completely change how your agent behaves without your knowledge or consent. Not to mention that you might have to go through and change how many connected "parent" agents work if the connected agent suddenly requires different inputs/outputs to operate.
Best practice is to use Component Collections and introduce child agents, knowledge or even topics to your "parent" agent as managed solutions. This way everything is versions and no one can pull the rug out from under your agent. You decide when it get's the update and deal with any changes through dev/test/prod ALM.

[adilei]

@inescavem-ai , agent outputs will automatically get picked up by the Orchestrator. If you need to store outputs in variables, use a topic with inputs, and instruct the orchestrator to chain the connected agent with the topic.

[adilei]

@johnsonmattMS, Dewain’s passion on this topic has a way of guiding us all 🙂

There are some risks in taking dependencies on external agents (MCS or otherwise), but I wouldn’t go so far as to call it best practice to avoid them altogether — sometimes those dependencies are exactly what’s needed. That said, the potential risks aren’t the reason inputs and outputs are YAML-only for now.

[johnsonmattMS]

@adilei I did laugh at myself after I posted my comment then read who wrote the original article. Talk about trying to teach granny how to suck eggs. Haha.

[lmackenzie94]

Hi there, I've figured out how to edit the YAML of child agents in Copilot Studio but I'm not sure how to do it for connected agents. Is it not possible using the Copilot Studio UI?

Thank you!

[adilei]

Hi there, I've figured out how to edit the YAML of child agents in Copilot Studio but I'm not sure how to do it for connected agents. Is it not possible using the Copilot Studio UI?

Thank you!

Navigate to Agents -> You Connected Agent -> Switch to Code View

[tboswell-lbmc]

We added inputs to a connected agent today and it caused some odd behavior. The main agent inconsistently uses the input descriptions and asks for the inputs in random orders. Also, it treats any feedback from the connected agent as completion of the task. To elaborate, we have a connected agent that’s creating a record in Dataverse, when the connected asks a clarifying question about the record the user is trying to create, the main agent responds saying “the record was created”. This only seems to happen when we define inputs, as soon as we remove them from the YAML, it goes back to relaying the follow up questions and completing the task. Are we doing something wrong?