Inter Processs Communication

[Leon0402]

Hi,

I have the following setup:

• An Code Writer Agent
• An Code Executor Agent

I want the code writer agent to be able to write code that can ask other LLMs. Imagine something like this:

def fact_checker(query: st) -> bool:
     model_client = OpenAIClientProvider.client.get_instance()
     response = await model_client.create( ....)
     return response.content.answer # Assuming some structured output here

It is a little bit simplistic here, but the basic idea is that an LLM can be called from within code and the result is available as a variable that can be directly used in the same code snippet.
The issue is that these calls the OpenAI (in the future it could also be a full autogen chat) are decoupled from my main process. So nothing is logged for instance.

It would be really cool if I could integrate it somehow in my main setup such that its part of the inner messages of code executor agent for instance. Open for ideas here.

Currently my approach would be probably to save the output of this agent in some file and then try to parse that again in the code executor agent. But I am open for more ideas!
I also heard autogen supports some tracing solutions. So maybe that could also help here? I am not too familiar with it.

[ekzhu]

There is a user defined function feature for command line code executor.

It's https://microsoft.github.io/autogen/stable/reference/python/autogen_ext.code_executors.docker.html#autogen_ext.code_executors.docker.DockerCommandLineCodeExecutor

See the functions parameter.

It allows you to have access to predefined modules and functions inside the code executor. Maybe you can write your "sub process" workflow as a function.

For tracing, see https://microsoft.github.io/autogen/stable/user-guide/core-user-guide/framework/telemetry.html

[ekzhu]

The function feature is a bit experimental. So let us know anything can be improved.

[Leon0402]

@ekzhu But this is only for passing function / module definitions in it, right? So I cannot pass in sort of instantiated function that use my existing python object. In that case, I already have that part implemented. So I use Jupyter Code Executor and I basically execute in the beginning imports and inject some data (copy - So I basically generate code cells dynamically that are executed within the jupyter notebook, I cannot inject like actual objects as the jupyter notebook of course executes in its own process).

So my request was more about the "communication" part between the main process and the jupyter notebook / code execution agent. I could implement something self made, but I was hoping that autogen has something useful built in that could help with the implementation.

[Leon0402]

Regarding tracing: Is there any doc on how to do it in agent-chat. Tracing seems to be built in the core framework with SingleThreadRuntime. I found a little bit documentation on how to use for instance Agents in autogen-core. But if possible I would like to use agent-chat as much as possible, so I wonder if there is any guide how I can do some more low level stuff like tracing without getting rid of RoundRobinChat, AssistantAgent and all these nice things

[ekzhu]

I see. I don't have a great idea for this one at the moment. It would be great if we can discuss more either during our office hour and on discord.

For tracing agentchat: this will become possible once we make the runtime customization available at agent chat layer. Say 2 weeks.