Python : Sample for Swarm Agent Approach

[nmoeller]

Motivation and Context

In OpenAI Swarm & Autogen Swarm Agents are able to handoff via Function Calling, rather than using a Prompt in the Selection Strategy. This PR should provide a sample how you can build a Swarm Agent System in Semantic Kernel.

The example implents the sample of Autogen Swarm.

Contribution Checklist

• The code builds clean without any errors or warnings
• The PR follows the SK Contribution Guidelines and the pre-submission formatting script raises no violations
• All unit tests pass, and I have added new tests where possible
• I didn't break anyone 😄

[markwallace-microsoft]

Python Test Coverage Report •

File	Stmts	Miss	Cover	Missing
semantic_kernel/agents/strategies/selection
swarm_selection_strategy.py	29	29	0%	3–71
TOTAL	19047	2409	87%

Python Unit Test Overview

Tests	Skipped	Failures	Errors	Time
3199	5 💤	0 ❌	0 🔥	1m 41s ⏱️

[nmoeller]

Hey,

so this is just a sample how an Agent handoff could be done with Function Calling.
Would you be interested in this integrated in Semantic Kernel Agents ?

I think i could also make a class SwarmAgentChat,SwarmSelectionStrategy in SK.

Just let me know if you would be interested in a more sophisticated implementation or the Sample is enough.

[moonbox3]

Thanks for working on this @nmoeller. It looks like with OpenAI's release today of their Agent SDK, Swarm is no longer going to be maintained. With that said, they introduced the "handoff" feature in the new SDK. We can have some equivalent behavior in SK with something like this. What are your thoughts?

[nmoeller]

@moonbox3 Thanks for looking into this. Let's probably first discuss if my sample still makes sense compared to the one @sphenry provided. What i wanted to show in my sample is how to handover with KernelFunction / function_calling , I think this is definitely included in the sample aswell.

I have two thoughts on this :

1. If providing the history as parameter for the function does the LLM determines the history or is the History injected via the KernelArguments ? I think from a performance perspective maybe it's not good to have this parameter determined by the language model.

2. The sample doesn't do multiple handoffs in a single Iteration, this can also be implemented with the strategy he's chosen but i think this would be easier with using a AgentGroupChat. But the advantage here is definitely it is more readable than mine and does not need so much pre-knowledge of SK.

Conclusion :
From my perspective i think both samples make sense in the repo, One with more autonomy the other one with less autonomy. But we should align the samples.

I would adjust my Sample to be closer to his sample proposing the following changes:

1. Removing the Swarm naming and just calling it a "handoff pattern via KernelFunctions with Agents"
2. I would refactor my code to be closer to his Handoff Plugin

What do you think ?

If you think it does not make sense we can also close this PR.

[nmoeller]

I will close this due to #11184 , which brings a similar approach, if this is still needed feel free to ping me, then i will reopen.