recursive agent harnessing (Harness²)

[samuelscheit]

As google showcases in their Harness² research project,
the performance of the model can greatly be improved by letting the agent engineer and improve it's own harness.

T3Code could adapt that approach, tho I don't know about any specifics and opened this to initiate the discussion

might be related to #17253 #6741 and #10242 (tho only hardcoded orchestration capability improvement)