Long running agent loop like code assistants?

[bbence84]

I have a question whether Semantic Kernel is actually suitable for a kind of long running agent loop that some use cases might require. As an example, you provide an instruction to a code assistant (like Github Copilot agent mode), provide a context (list of source code files and additional resources), and the agent starts to interatively work on the task. It might face errors occasionally, which it can usually recover from. So it's like a long-running agent loop. Not sure if this kind of orchestration and agent loop is supported by Semantic Kernel. So far, I have been using Semantic Kernel that makes use of the legacy Handlebars planners and more recently, the native OpenAI function calling "planner". But such use cases like a coding agent might require some additional orchestration logic to make it to work.
Thanks!

[moonbox3]

Hi @bbence84, have you had a look at our new agent orchestration patterns? IIRC, you work with Python, correct? I would take a look at our "magentic" pattern, inspired by AutoGen's "Magentic-One". The magentic patterns works based on a plan, and can attempt some "course-correction" while working on the sub-tasks.

For more advanced scenarios, one can have a mix of the agent orchestration patterns along with our process framework, which allows for an event-based (directed graph) orchestration flow. This is when you not only want to rely on prompting, or unstructured agent tasks via the orchestration patterns/agent framework, but want to make sure step A always leads to step B followed by fanning out to step C/D and then fanning in to step E for the final result (solely as an example). We have some "getting started with processes" samples here along with some other concept samples here.

Have a look, and if you play around with them, we'd love your feedback. Thanks.

[bbence84]

Thanks a lot!