AutoAgent turns harness tuning into self-improving loop

The real story is not "agents got smarter overnight"; it is that the harness became the thing being optimized.

• –Strongest angle: this reframes agent quality as an engineering problem around tools, prompts, workflows, and eval loops.
• –The domain-specific framing matters: if the feedback loop is good, the same system can be tuned for coding, spreadsheets, research, or other narrow tasks.
• –The benchmark claims are compelling, but they are still benchmark claims; I would treat them as evidence of direction, not proof of broad generalization.
• –The "same model evaluates the agent" idea is interesting because it can tighten failure analysis, but it also raises the usual risk of self-reinforcing eval bias.