Covers what goes wrong once an autonomous AI agent can act, not just talk — destructive commands on real targets, over-broad permissions, retry spirals, infinite loops, and cheating to hit an exit condition. Use when an agent holds shell/filesystem/DB tools,…
Idioma del texto original: inglés