how i set up 12+ claude 4 code agents to work in parallel mostly autonomously for several hours without conflicts or slop code

in the video i uploaded yesterday the claude code agents were all working on building out eigencode infrastructure to make them even more capable, basically i used claude code to build the bridge to eigencode to then work on eigencode to make claude code more powerful to work on eigencode - creating a recursive improvement loop.

Want a strategy that stands the test of time? Learn how Walk-Forward Validation can stress-test your models across multiple market conditions. Full code on my Substack:

the slop code claim is sus

If you are doing this type of inference, you should definitely buy your own GPUs and self-host a model like QwenLong-L1-32B

thats the plan

How do you quality control this? What happens if they go into a rabbit hole and you have to step in? I guess at this point you are no longer familiar with the code they wrote and you are basically managing them.

Thanks for sharing so much Thermo. In my world of AIassisted coding, I let the agent (Windsurf Claude 3.7 primarily) work on the entire repo. My more ambitious projects are a collection of services, databases, and front-ends. And I avoid libraries as much as possible, coding at the bare metal as much as possible. In my style, I rarely am editing just one file at a time (for example Typescript). So, could you see multiple agents taking turns on a repo? Before one pauses, it documents its progress and prompts the next one to continue.

yeah - i've already been doing that with taking turns and documenting progress when stopping for a few months - this is a step beyond that where i have multiple agents share context and update progress with each other while working

This makes no sense dude. Even with file locking the agents are going to be working against each other. The only way I see this working is if the app itself building is microservice based and each agent has its own domain of stories to work on.

Show results