Single vs multiple instances of Kernel Object

[smokhasi]

We are using an agent group chat with AzureAssistantAgent and generally create a kernel object per agent in the agent group chat. This allows us to add the specific tools required by the agent to their own kernel object.
However, I was curious if this is the right way or should we be sharing a kernel object across all agents in the group chat?
How is the chat history managed in each case?

    database_agent_kernel = Kernel()
    database_agent_kernel.add_plugin(plugin_name='sqlite', parent_directory=plugins_directory)
    
    research_agent_kernel = Kernel()
    research_agent_kernel.add_plugin(plugin_name='serper', parent_directory=plugins_directory)

    chat = AgentGroupChat(
        agents=[database_agent, research_agent],
    )

[sophialagerkranspandey]

Adding @alliscode

[moonbox3]

@smokhasi, that's a fine way to encapsulate plugins for a specific agent. Doing it that way ensures that that particular agent only has access to those specific tools. The other possible way to manage it, if you did need only one kernel instance (again not saying you must do it this way, it's an option), is to then control the available tools via the FunctionChoiceBehavior, which allows one to configure a filters dictionary either including or excluding plugins. This is more work, but again, and option. If it were me, I would stick with separate kernels per agent.

Regarding chat history: context is broadcast to each distinct agent type via the agent channels. Each agent, right now, has the full context of the conversation.

[smokhasi]

Thank you, @moonbox3 .I will continue to use a separate instance of the kernel per agent.
Regarding chat history, could you please elaborate on strategies to pass the right context to the right agent, instead of passing the entire context to all agents?
We have found that agents tend to provide non-deterministic results with larger contexts.

Right now, as part of individual agent persona, I have been instructing the agents to read the chat history and identify the required context. Ideally, I would like to avoid providing that instruction via a system prompt and instead pass only the specific context required by that agent.

[smokhasi]

@moonbox3 additional question on using different kernel objects per agent is how do I control the selection strategy in this case?

chat = AgentGroupChat(
    agents=[agent_reviewer, agent_writer],
    selection_strategy=KernelFunctionSelectionStrategy(
        initial_agent=agent_reviewer,
        function=selection_function,
        kernel=kernel,
        result_parser=lambda result: str(result.value[0]).strip() if result.value[0] is not None else WRITER_NAME,
        history_variable_name="lastmessage",
        history_reducer=history_reducer,
    ),

If I were to create a kernel object per agent, which kernel object needs to be passed to the KernelFunctionSelectionStrategy?

[moonbox3]

This kernel object for either the KernelFunctionSelectionStrategy or KernelFunctionTerminationStrategy can be anything. It really only needs an AI service registered so that you can get a model response back to help determine the selection or termination strategy. So that particular kernel is pretty slim.