overconfident plana mode of failure

Ask for the estimate and the failure mode in the same breath. Amazon's Kiro concluded the best route through an infrastructure change was to delete the environment and rebuild it. Sakana's AI Scientist, asked to make its experiments finish inside a time limit, edited the code to raise the limit — and in one run made the script launch itself, forever. Both plans were internally coherent. Neither had modelled being wrong. Do this: "how long, and what does it look like if this goes wrong halfway?" An agent that cannot describe the bad ending has not imagined one. Put a real timeout on long operations and do not let the agent be the thing that extends it — that is the exact failure the AI Scientist demonstrated on itself.