Can we invoke Agent through kernel?

[HuskyDanny]

We can invoke function or plugin with the kernel, so the code pattern was to pass around the kernel so that the downstream can get whatever tasks they want, but can we do it for agent? I cannot find a way to get an agent or invoke agent through kernel.

[rogeriorfp]

@HuskyDanny Maybe this helps: Step08_AgentAsKernelFunction

[moonbox3]

@HuskyDanny, as of SK Python 1.27.0, we've provided a way to create a plugin from an agent. We're showing it in the following pattern, but it would hold to add an agent as a plugin, and invoke it from a different way, i.e. kernel.invoke(...).

https://github.com/microsoft/semantic-kernel/blob/main/python/samples/concepts/agents/chat_completion_agent/chat_completion_agent_as_kernel_function.py

[HuskyDanny]

Thanks @moonbox3 , do we have example on how to invoke the agent through kernel? I found we might call the model_post_init to make it a kernel function. so is the way the recommended?

agent = ChatCompletionAgent(
            kernel=kernel,
            name="Nl2sqlAgent",
            prompt_template_config=prompt_config,
            arguments=arguments,
        )

        # Make agent as a kernel function
        agent.model_post_init(None)

kernel.invoke("Nl2sqlAgent")

[moonbox3]

Although the agent can be run as a Plugin/KernelFunction, it won't work directly when invoking it via the kernel. You'd need to create a prompt function, add the agent to the kernel like kernel.add_plugin(my_agent), and enable auto function calling, which could call into the Agent.

The Agent has underlying methods that return an AsyncGenerator, versus a normal Kernel function that returns a coroutine.

[HuskyDanny]

Yeah I have noticed the pattern, I could invoke the agent through kernel, but found some issues as outlined here: #11638