I noted that the previous AI agent harness is now considered dead and a new harness is emerging.
I predict that in six months about 90 % of agent work will run off a user’s laptop.
The cost is real.
Self tuning harnesses fix it.
Remedy workflow: Swap the model → rerun the loop → done.
I concluded that using self‑tuning harnesses with a simple swap‑and‑rerun sequence resolves the genuine cost issue.