@ColeMedin - Archon V7 will have LangGraph Documentation. I have suggested you include CopilotKit documentation. Over time, users will want integration of other documentation.
Instead of you creating the documentation integrations one by one, would it be possible to create a feature that allows users to give Archon the documentation URLs for any frameworks, tools, or APIs they want supported, then have Archon automatically create the documentation integration?
If so, you could create another feature that allows Archon to access a database of documentation integrations created by other users.
CopilotKit is on my radar! I appreciate the suggestion
Also yes, I would love to eventually get to the point where users can choose the documentation Archon uses instead of selecting from a predefined set of frameworks!
The biggest challenge with that is if I use a predefined set of frameworks, then I have the opportunity to optimize the prompting for these frameworks and support issues people have with them better. But it’ll very much become the wild west once people start trying Archon with frameworks I haven’t optimized it for! Still will be an awesome feature though.
Yeah, optimisation was the bit I wasn’t sure about. Namely, to what degree can it be systematised and automated? I take it the optimisations are mostly in the form of system prompts? As models get smarter, perhaps they could assist with this? Even having agents that initially got things halfway there, via knowledge of documentation and common technical problems, would be useful. Then there could be ongoing improvements with things like automated testing, hallucination detection, automated error detection, human in the loop feedback, and manual tweaking. That said, I feel for you regarding the support implications of including frameworks you did not personally integrate. Even if automated or semi-automated optimisations worked well, you would not always have familiarity with the framework’s workflows, technical requirements, and documentation; thereby making it difficult to offer support. If this feature was viable, maybe just offer it with the caveat that support for unofficial integrations would be limited, much in the same way that CMS projects don’t officially support third party plugins. Also, if a feature like this was reasonably effective, it could help you to more rapidly offer official integrations.
Yeah you definitely get me! Correct - a lot would be through the system prompts, but also potentially the framework would dictate the different subagents I create to handle different parts of the agent creation.
Like with Pydantic AI I would create an agent specifically to define the dependencies. But with CrewAI I probably wouldn’t.
Maybe a way to generalize this and LLMs can assist with the system prompting for sure!