Extending this project

[Mergpijp]

Hi,

Thanks so much for your contribution and making it understand. I understand you want to make this mechanical without magical feelings but one can't resist they are LLM's doing some leverage in a sense. I am interested in mistral as its open-weights and very good licensed.
For production you recommend using state of the art models from anthropic for example? I would wish to scale it make it more human friendly with more content. Do you think its production ready if you have a good gpu and run it on this together with your wonderful ideas on how to make it? What would you propose come after this? Are you planning this?

Kind regards,

Emmanuel

[pguso]

Hi Emmanuel,

Thank you for your thoughtful message, I really appreciate it.

Regarding production use: I do believe that several open-source LLMs are already capable of being used in production, provided that the proper safeguards and control mechanisms are in place. In my view, this is true regardless of whether you use open-weights models or closed models like OpenAI or Anthropic. Guardrails, validation layers, structured outputs, monitoring, and fallback strategies are essential in both cases.

Personally, I particularly like the Qwen 3 models for certain agent tasks. I’ve found them to be very strong in structured reasoning and tool-oriented workflows, which makes them well suited for agent-style architectures. There are also quite a few other OSS models that I tested and loved.

I’ve also experimented quite a bit with deployment setups. So far, I really enjoy using modal.com. It’s very developer-friendly, allows you to deploy OSS models with very short cold starts, and you can only pay per invocation instead of paying for idle GPU hours all day with the right setup. That makes experimentation and scaling much more practical compared to running dedicated GPU instances continuously. I thought about publishing a website where devs in three clicks can deploy some open source models to modal and have a api endpoint ready in about 30 seconds, but I don't know if the interest is there for this.

As for extending this repository, I would absolutely love to expand it further. What’s currently missing for me is clearer feedback from developers about what they actually need next. For example:

• Better production patterns?
• Observability and logging examples?
• More structured tool integration?
• UI layer examples?
• Benchmarks or performance comparisons?
• ...

If I better understand what developers are struggling with or wishing for, I’d be happy to evolve this project in that direction.

Thanks again for your interest and I’d be very curious to hear what specific direction you would personally like to see it go.

Cheers,
Patric

[Mergpijp]

Hi Patric,

Thank you for your response. I was thinking of making this a custom agent producer, something that is kind of a factory and grows itself according to specification. I notice also that there can still be some improvement in making the lesson better integrated for example atomic_action which failed with me returning None and now does a good job can be used in lesson 10 atom_of_thought. I am moving a bit away from lessons but more towards a holistic project. You already do a very good job with the agent state being extended. I was planning to build a pdf reader agent for my project to get necessary data but it quickly becomes very specific thats why an agent factory isnt a bad idea imo. Also your other repo about RAG can be helpful to integrate into this holistic project. It's a lot of work but fun to do especially since you have laid very good foundations to this agent tutorial. For me its an exercise in keeping the code base clean and re-usable. Let me know what you think.

Kind regards,

Emmanuel