Every agent can complete any microtask successfully. At the same time, every agent will fail once complexity of scope and context gets too high, and it has to navigate planning, implementation and review on different abstraction levels. We realized this and built zeroshot, where planning, implementation and feedback is delegated to isolated agents. The feedback agents have non-negotiable veto powers over the worker agent, meaning that the usual tendency of optimizing for completeness over correctness gets avoided. We are using zeroshot to build a 1 million LOC hosted platform, and due to our trust in the tool to produce senior-level, scalable code, we have completely stopped looking at the codebase.

I agree. Can you say "1950s supercomputers" vs modern day cell phones? I am curious how you manage consensus of feedback agents or is it 1:1?