A new paper says stop retraining the model, fix the harness

There's a quiet thesis building across AI research this year: the model is mostly done improving on its own, and the gains now come from the harness around it. A new paper called Adaptive Auto-Harness pushes that idea to its logical end. Instead of a static agent that's frozen the moment training stops, it proposes a system that keeps improving itself while deployed, learning from an open-ended stream of real tasks as they come in.

The mechanism is the interesting part. The agent runs on whatever tasks show up, watches how it does, and adapts its own strategies and scaffolding, the prompts, the tools, the routing, to sustain performance as the task distribution drifts. No retraining the weights, no waiting for the next model. The harness is the thing that learns. It's on arXiv as 2606.01770.

Read it next to the Berkeley scale-the-system-not-the-model paper and Ant Group's SkillAdaptor from the last couple weeks and the pattern is unmistakable. The center of gravity in agent research has moved off the model and onto the scaffold around it. If that holds, the moat for an agent product stops being which base model you use and becomes how well your harness teaches itself on your users' actual work. That's a very different game.