How to implement a multi-agent "chat" workflow using the Semantic Kernel Agent Framework?

All answers to this question.

The Agent Framework in Semantic Kernel is specifically designed for these multi-turn interactions. You define each agent with its own persona and plugin set. To handle the coordination, you use an AgentGroupChat. You can set a TerminationStrategy—for example, stop when the Reviewer Agent outputs "APPROVED" or after a maximum of 5 rounds. The framework handles the passing of the message history between them automatically. This is much more organized than trying to manage two separate chat instances manually, and it makes it very easy to add a "Human-in-the-loop" step where a person has to approve the final output before it's sent.

   Answered 2025-12-01 by Natalie Rushman

Can these agents run in parallel, or is the hand-off strictly sequential? For example, could three "Researcher Agents" work at once and then report back to a "Summary Agent"?

   Answered 2025-12-05 by Tony Stark

• Currently, the standard group chat is sequential, but you can definitely implement parallel execution using standard C# Task.WhenAll. You’d have the researchers run their tasks concurrently and then feed their combined outputs into the Summary Agent's kernel. It gives you a lot of flexibility since it's just standard .NET code at the end of the day.

     Commented 2025-12-08 by Steve Rogers

The best part is the state management. You can pause the chat, save the state to a database, and resume it days later when the human reviewer is ready.

   Answered 2025-12-12 by Wanda Maximoff

• This "persistence" is exactly what was missing from our earlier prototypes. It makes long-running business processes much more feasible with AI.

     Commented 2025-12-15 by Peter Parker