I like the framing of "harness problems" vs. "model problems." A lot of teams jump to fine-tuning when the real bottleneck is that the agent lacks the right context, tooling, or feedback loop. Treating execution traces and artifacts as evidence, then validating harness changes before promoting them, feels much closer to how we improve traditional software systems.

The emphasis on regression testing candidate harness changes also stood out. Without that guardrail, it's easy to optimize for one workflow while quietly degrading five others. It'll be interesting to see whether Harneloop generalizes beyond Blender into areas like coding agents, browser automation, and enterprise workflows, where the harness often has a bigger impact than the underlying model.