Context Engineering for Agents

I call it feature engineering for agents (arxiv), similar to what people used to do in classical ML to improve the input representation. Once you have a good representation of your data (by selecting, merging, mapping, etc.), even a light model can surpass a complex models (e.g, Logistic->xgboost). For agents, these features are the context. So, it is not only the goal (pilot using big LLM), it is the system (components besides the LLM). The idea is to develop an agent/workflow/app with the smallest feasible LLM so that you are forced to improve the system (cognitive architecture) first. Of course, LM can be replaced by any other engine.

This is such an amazing resource!! I appreciate you compiling so many great supplementary content

Nailed it, LangChain! Context engineering strategies are top-notch. Can't wait to see LangGraph in action.

Great Refresher Video On Context Engineering, Perfect for Sunday Morning!

Thanks for sharing the knowledge, enlightening.

really amazed how context management plays a vital role in building a good agent rather than using a LLM with large params

really cool stuff harrison! we use state object to manage state so super simple and intuitive to pass along only the required stuff to child graphs even fir example.

this is super interesting. the concept of trying to achieve a procedure with as small of an LLM as feasible by optimizing context window is certainly a very useful one. just yesterday an open source opanAI llm model came out that works locally on my 9070xt and i've been copying and pasting small contexts (it can only hold a very small context window before being overly saturated) and then iterating to try to identify how (at least one small) model interprets context. i think having this stuff available on youtube is huge because we need to be transparent about work and development in this cutting edge stuff to optimize use. one thing i've realized from my limited experience is that a hallucination of an llm can be leveraged using recursion to enhance a model's performance. I'm still in my infancy of my development and learning python before i get into more advanced things like RAG and proper context (prompt) engineering (on coursera) but this was a beautiful glimpse!

Another great video! Thanks for taking the time to make these

Amazing video on organizing context engineering assembling such great context content in this video. Thx!