Reflection Agent From Scratch | Agentic Patterns Series

[@franklinobasy]

Great Explanation! Welldone.

[@TheNeuralMaze]

Glad you liked it!

[@abhijithkadam383]

Great explanation, thank you

[@TheNeuralMaze]

thanks! 🙏

[@vq8gef32]

Amazing and thanks for sharing this

[@TheNeuralMaze]

Anytime! Thanks for the kind words 😍 I'm already working on the second episode! 🦾

[@Danyal_alam]

Hey Miguel, can you please create a tutorial on production ready recommender system?

[@TheNeuralMaze]

hey! Once I finish the "Agentic patterns from scratch" series yes! that's what I had in mind - Medium Posts + KZbin Videos about production-ready RecSys 😃

[@paul_devos]

Great content, thank you! Question for you, what tool did you use to draw your diagrams, those are rich and love the fonts you used.

[@TheNeuralMaze]

thanks!! It's Excalidraw 🙂

[@ThisIsChina-mu1ls]

In my ReAct multiple agent system, the time it takes to obtain the final answer using tools can be quite long, leading to a poor user experience. In your system, how long does the process of calling the choose tool -> execute tool -> get final answer take in each round,, and how many times does the LLM (Large Language Model) need to be called?

[@TheNeuralMaze]

It depends on the task, if you go to the third video of the series, you'll see the ReAct agent I implemented from scratch, using three tools. The problem it has to solve is very simple, but maybe it will give you an overall idea of the time it takes.

[@nessrinetrabelsi8581]

When you loop on (generate then reflect), you basically are not taking into account the last reflection (critique) of last loop, no?

[@TheNeuralMaze]

Hey! So previously I was taking into account the whole history of messages. Problem? You may end up with too many tokens for the LLM window. What I did at the end was to limit the chat history to just two messages (the previous reflection and generation completions), and the results were very similar (and much faster).

[@web3global]

🚀

[@TheNeuralMaze]

🔥🔥🔥