An article collecting coding-agent failures in real repositories discusses unverified claims that work is complete, conflicting instructions, and retries of actions with uncertain outcomes. The accessible introductory text recommends retaining check commands and results as completion evidence, organizing instructions, and verifying idempotency. Separately reviewed Codex documentation distinguishes sandboxes that limit what can run from policies governing when approval is needed. The cited original was inaccessible, so only the publicly available explanation, not all fifteen cases, was verified.