Here's where I was hoping openly available models would shine. Some community gets together, starts sharing successful/failed runs with their own agent, start building a open dataset for their specific syntax and tooling. then finally finetune new variants with it for the community.