Complex orchestration scenarios

[Ilia-Kosenkov]

Hey everyone! I am exploring SemanticKernel as a replacement to our ad-hoc implementation of agentic RAG. While our implementation is a bit limited and quite crude, the Agents/Orchestration framework in SK seems to be quite flexible and can greatly improve our service.

The problem I am facing is that I seem to require support for child orchestrations. Let me explain the scenario:

• We start with a user prompt and a classifier agent, which can decide, if the question is relevant, can be answered from history, or should be directed into one of the two 'subworkflows'. One workflow is built around semantic search and is represented by a singular agent. Another workflow is a complex multi-step data retrieval, which consists of 3 agents.
• Right now, I managed to achieve adequate results by utilizing simple handoffs: start with supervisor and delegate to one of the workflows if needed. However, for the complex workflow, I need to hand off not to a single agent but to another orchestration (in an obscure manner, right now I use sequential orchestration for subworkflow, but tomorrow I might go magentic or group chat or whatever).

Using durable functions for analogy

If you are familiar with durable functions, then orchestration in terms of SK resembles Durable Orchestration, while each agent is an Activity. Right now we can easily spin up an Orchestration which calls Activities, while I am looking for an ability to also call SubOrchestrators.

Essentially, my question is, how bad would it be if I create an agent, which itself is an orchestrator? Are there any similar solutions already available (docs and samples are sometimes too broad or generic)? Maybe there is a better solution, where I can derive a custom orchestration with unequal transition rules (e.g. when agents become visible only if certain condition is met, etc) or as a combination of other orchestrations?.

Thanks in advance!

[arafattehsin]

Hey @Ilia-Kosenkov - Whilst this can also be achieved (I think) from Agent Orchestration but if I am creating it from the scratch, I would always look at the combination of the Process Framework and Agent Orchestration.

https://github.com/microsoft/semantic-kernel/tree/main/dotnet/samples/GettingStartedWithProcesses/Step04 <-- Check this out!

[moonbox3]

This is a good start, @Ilia-Kosenkov. Let us know if you have further questions while looking at the process framework. Also take a look at our Magentic orchestration pattern: https://github.com/microsoft/semantic-kernel/blob/main/dotnet/samples/GettingStartedWithAgents/Orchestration/Step05_Magentic.cs. If you need Python links, please let me know.

[Ilia-Kosenkov]

Thanks! Process Framework is something that I have not touched yet, perhaps it is the way to go. I used Magentic orchestration but the issue with it is significant overheads -- it makes multiple calls to chat completion endpoints before invoking an agent, which also slows down the whole thing. The reason I am trying to invent something flexible here is that if I can provide an answer before spinning a magentic orchestration in se scenarios, it would be much better from the UX perspective.