Hamro Channel

i have a VERY VERY stupid simple way i am doing this, and will likely get made fun of for how I am doing it... but I gave each of the 4 systems their own physical box (tho 2 are VM's, but each VM is on its own Mac host and the VMs run ubuntu arm). They all have ssh + sudo access to each others machines. I have shared network drives linking all machines together, and cronjobs + shell scripts monitoring for 'f;ags' to be set. When Claude (my manager) wants to assign a task to one of the other models, it opens up a Assignments.md file in the shared drive, and puts who its assigned to, what the job is, what it needs to produce, what agents it should use to complete the tasks and what LLM's should be reviewing its work. It then creates the 'flag' (this is just a file in the shared drive named message_for_gemini as an example). gemini's cron job will detect that flag within 60 seconds, and will then run a shell script that invokes gemini in yolo mode, its passed a very small system prompt that tells it how to parse the assignments.md file, and how to use my shared_comms.md file (this one is basically a place for them to 'chat' and ask each other questions easily and quickly). once it parses the assignment, it will launch the agents needed to perform the task, and once it has something it thinks is ready, it will create the flag for whatever LLMK is suppose to check its work. if they both agree, it sends the info to Claude with the shared_comms and message_for_claude flag creation. claude then reviews it. if claude thinks its wrong, it will either ask me (if it has evidence on why its wrong) or will ask the 4th model (most of the time grok) for its input. if grok agrees its all good, then claude accepts it and keeps the process going. if grok or claude think its wrong, they send me a text message asking me to check the laptop to provide feedback and a final decision. it seems confusing af writing it all down, but its been working really well for about 2 months now. They also all have to keep a running log of every single command/tool they run (with all params) and some other stuff, depending on the task at hand (and whether its something i want to learn, or already know how to do).