The productivity gains of people using their own "custom" harnesses and agent tooling (e.g. letting agents talk to eachother over a messaging board) is questionable at best a waste of time at worst. No one I've talked to is actually applying the scientific method and measuring how long and how many tokens it costs to get a certain task done with just a plain chat input vs all this custom harness/tooling crap.

All you need is the source code for your existing project and the claude code/codex/opencode cli and boom, you are ready to ship. Most people doing the impressive stuff with AI are just using it, not building harnesses or "agentic" tooling.