Context Engineering is the future of AI Agents - here’s why

Admittedly I'm only at the 7 min mark of your video, but what you're displaying there is nothing but a classic "trivial parallelisation" of tasks with inherent dependencies between them. This is nothing new at all, and should actually be blindingly obvious - it doesn't need a full research paper to explain it. Try this sometime...get 3 people and give them the task of each generating a number, that sum to 100. But they have to each stand in a different room, unable to interact with eachother while doing so. How do you think that is going to play out? It seems you are conflating the concepts of "agents" and "workflows" (in my humble opinion) - You can have as many agents as you want, with each one specialising in something specific - you dont need a "linear agent" at all - not one little bit. You need a "linear workflow" that ensures that no parallelisation of tasks with interdependencies are carried out.

This method will quickly overwhelm the context window when applied to the complexities of real-world problems. This is why my architecture uses workflows instead of agents, and why I use a multi-layered memory (state) management system which is hardcoded into my architecture. I have agentic workflows which manage the memory spaces of the system and its components so that only information relevant to a specific task is given to the context window of that specific agent within the workflow. This micro-management of memory states according to task really makes a difference in the reliability of my MAS.

LangGraph - tool based handoff - carries state with it.

It reminds me of something I've already encountered in the Novelcrafter program, which uses AI to write stories. But they did something more there. When you write a new scene in a story, the AI also includes summaries of previous scenes in your instructions, allowing the narrative to be maintained. However, in addition, the program automatically includes in the prompt descriptions of important people, places, and objects that appear in your instructions, so that the LLM writing the next scene has this information. Example: your instruction: "The next scene describes a quarrel between Tom and Jessica". The program will add summaries of previous scenes and descriptions of the characters - Tom and Jessica, to this instruction. The program contains a database of key characters, places, and objects related to the story.

Honestly, this just makes sense. Whenever I design a system, I typically reach for paralellism last. Paralellism is great for getting large tasks accomplished more quickly, but you need to be sure that all the necessary input is available (in this case, the "context") and that the operations can be run in isolation from other processes/workflows without them requiring or impacting one another.

🤖Join the New Society and get free Vectal Pro for one month- Go here: https://www.skool.com/new-society ✨July only offer!

Uhh regarding parallel agents in Claude code im always running them in parallel for research and it works fine and for refactoring tasks that dont conflict.. it works well ..

I really liked the intellectual drive to admit mistakes, even in past videos. It's clear that you're a serious guy, that you're smart, that you work well, that you love what you do and that you do it well.

Could not agree more. Ive hit a brick wall after working on a huge (for me) multi agent project over the last 6 months. Sharing context and an application level memory layered on top of the agent memory. Appreciate the vid man, youve made a few things clearer for me.

As always well, focused and condensly presented. Thanks for your effort and input 🎉 Good to have you.