I Configured My Agent to Agree With Me
A rival model's review of my agent's work looked sharper — it pushed back, it demanded alternatives. So I ran a controlled experiment on my own harness: same model, same prompt, config on and off. The vanilla arm argued with me. Mine never did.