Python: using agents as KernelFunctions

[eavanvalkenburg]

Motivation and Context

This adds a convenience method to allow a agent to be called by other agents or regular function calling.
Also includes some improvement in parsing of classes into KernelPlugins

Description

• Adds a function to a Agent that uses the get_response and a single string 'Task' to run.
• It takes the name and description of the agent for the function name and description and the task and output are annotated
• Adds a sample showing how to have hierarchical agents calling each other.
• Could be extended to also pass in chat_history, which currently doesn't happen.
• Also adds some extra features to plugin creation and adding, now static methods in a class also get picked up, instead of just bound methods
• Also checks if there is a attribute called 'name' and take that as the plugin_name if it exists.

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/functions
kernel_plugin.py	181	6	97%	393, 396, 409, 434, 455, 480
TOTAL	19619	2561	87%

Python Unit Test Overview

Tests	Skipped	Failures	Errors	Time
3211	5 💤	0 ❌	0 🔥	1m 40s ⏱️

[TaoChenOSU]

While I think it's ok to have agents as tools/plugins, I don't think that is a pattern we should encourage. We have the CrewAI as a plugin since it couldn't fit with our abstraction very well.

I am mainly concern about the following:

1. Agents are supposedly used for long running tasks, and they may modify their environment. This may not fit well with the previous chat completion tasks, i.e. simple questions & answers.
2. When should we advise people to invoke agents directly and when to use agents through kernel functions?
3. I am not sure if passing an agent to another agent will be a pattern we will eventually have. It may be, but I think it is worth some more discussion. With the shared agent runtime that is in planning, we may end up doing this hierarchical pattern differently.

[TaoChenOSU]

This is not going to work with other types of agents except the chat completion agent.

[eavanvalkenburg]

For my info, why not? I saw get_response is a abstract method so should be available on every agent, or are the args different?

[TaoChenOSU]

The args are different.

[moonbox3]

Superseded by #11184.