This is a GREAT framework

[jmm2020]

This is a great Framework, or more importantly IMO, guiderails for LLMs to be productive and focused over far lengthier projects than is normally possible. I would also venture to say given my usage an Expert UX/UI designer agent is also helpful. I have also been working on a Handoff coordinating agent as well that keeps the handoffs neat and orderly as well. Great stuff though and very promising.

[bmadcode]

Wow I would love to learn more about your handoff agent!

[jmm2020]

I am using a ML implementation of my own design and its using Context7 libraries to train so it has constant access to relevant coding data. I inject margins of error to the library data (I made a training sandbox with a UI) and am tuning the bias for the actor/critic that way since it started off throwing too much back(critic). It scared the hell out of me last night when I tasked it just for ha-has with creating a total plan for the development of a 120-unit apartment complex I am about to start constructing. It churned away for about 10-15 minutes and I got 96 pages of the most amazing data I think I have ever seen for such a project, Timelines, Code information including city, state and federal contacts with emails, phone numbers and websites, all up to date permit costs, expiration info, then schedules aligned with the timelines for the various inspections again pointing to the actual inspectors....I could go on and on I may have created a monster we shall see. I just sat there with my jaw dangling as I saw the file creation process. It also of course is doing QC to make sure everything is proceeding according to the architects' directives and throws stuff back(needs work). What really scares me is once I get this particular agent functioning "better", I need to fine tune a bit, I should be able to replicate the process with all the agents. The other cool thing I have been figuring out is I am fortunate to have 96GB of VRAM so I have been able to get multiple instances of LLM's performing different tasks(have had good luck with granite3.3:8b)...it uses one for negative and one for positive feedback, again working and apparently pretty well with what erupted out of it last night, and of course I have a reasoning model in there (claude 3.7) but after the first few steps it was called very little.

[bmadcode]

very cool - I am loving the integration of both context7 and also tavily, in combination they are great at finding the information they need!

[bmadcode]

also damn that is wild impressive - I need that hardware locally lol. I am running all local on an m4 macbook, no where near what you have - best i can do is spend cash on external llms such as in bedrock if I need that sort of power. Thanks for sharing though this is wild!