The agentic harness now sits above the model. It controls what models can do, remember, and access. This market is moving from isolated agent silos toward unified "meta-harness" layers that wrap, combine, and swap entire agent ecosystems into one cohesive framework. The open-vs-closed, owner-vs-renter fight that once defined the browser era restarts here.

Just this year, the labs proved the harness matters more than we initially thought. Welding their own harness to their own weights compressed a 21.8-point performance gap down to roughly 3 points at the top.

This month, Kimi K3 scored 88.3% on Terminal-Bench 2.1, just half a point behind Sol (Note: K3's training data is currently disputed. More on this in the report.)