I do not care about “local first” as a slogan.

I care about whether the system can make a practical call:

  • local when the loop needs to be fast
  • cloud when the task needs more headroom
  • and both inside one workflow without turning into a mess

That is the useful part of multiharness. It is a local-first agentic harness that can route between local and cloud models without hiding the path it took.

The DS4 piece matters for the same reason. antirez/ds4 is not interesting because it is a name. It is interesting because it proves the loop can be real enough that local inference is not just a benchmark flex.

My test is simple: can I see what changed, replay the path, and trust the failure mode? If yes, the harness is doing work. If no, the model choice is still theater.