Using Copilot I was able to write in Python a system that would parse YAML files containing prompts, formats, and dependencies. The code to do the topological sort was easy to do with one graph library. I didn’t add optimizations (like batching queries when possible) and it worked fine enough for me. Honest question: am I missing something or isn’t it just easier these days to roll your own integration?
True in some way, but don’t forget that most software is quite complex and typically developed by a team, not a single person. Does your integration allow users to easily customize it to their needs? And how do you handle edge cases? In larger projects, handling complex workflows, managing dependencies, and ensuring scalability often require more robust solutions, which is why many teams opt for frameworks that handle these challenges out of the box. Rolling your own is great for smaller, simpler projects, but complexity grows quickly in real-world application
@@codingcrashcourses8533 Thanks! That’s just the kind of answer I needed… indeed, speaking a common language by using a framework can serve as great documentation for the rest of the team (without anyone on the team writing documentation or having to update it). So, I guess the next question this leads to is: which framework is going to be additive without breaking backward compatibility. A dream framework can become a nightmare in three years if the “maintainers” keep “deprecating” behavior that was working but what they considered to be “too cluttered” or “technical debt.” I suppose one tell with that is seeing what dependencies the framework itself picks and what their own version history reveals. (This disturbing trend of “semver means we can break anything so long as we increment the right number” is something that extends far beyond just the Python community.)
Maybe i am wrong but arent you overcomplicating things in langgraph here? If you have one node that can use both tools you can skip adding the weather and airquality nodes and just add both tools to one llm calling node. Your supervisor node detetmines already what tool to call. Langgraph is in my opinion far superior then Swarm, which is experimental and unsupported. Let me know if i overlook anything here?
No, you are totally right, I could give one LLM all of the tools, but I could not think of a multi-agent approach that is somewhat simple, so I decided I create one LLM with one tool each to show how the LLMs are communicating.
You do great content, but these audio effects (quite loud) at every new bullet point killed the video. Suggest not over using it, and the levels are wayyy up
