Recommendations for best usage of current resources

Been reading papers on Magentic-One, Llama, Phi-4, etc. Really interested in the Magentic-One (Multi Agentic approach), and have some hardware to play around with. Please help me choose an ideal setup. Hardware that I have: 2x3090 1x2080Ti 1x970 (probably useless now) 1xK80 (also useless now) Computers: Intel i9-10900KF, 2x16GB DDR4, 2TB NVMe Ryzen 5700X, 4x8GB DDR4, 1TB NVMe, 500GB SSD (NAS) R730xd 2x12 Core E5-2678V3 (2.5GHz), 128GB DDR4, ~32TB HDD storage, 2x128GB SSD I am thinking I will put the 2x3090 in the intel machine, with NVLink, and try to run the 70b models in 4bit. I can use the 2080Ti in the AMD machine, running an 11B model. Overall, my goal is to fork Magentic-One, allowing for individually configurable agents with different LLMs. So if you were in my shoes, what models would you choose, and how would you leverage this? Right now I don't see myself training much more than a LoRA, and my goal is to have an LLM system capable of Software Project planning, code/repo surfing, and some code generation. Finally, what would your growth plan be after this? Move towards a single machine and more cards?