CrabTrap: I Put an LLM-as-a-Judge Proxy in Front of My Production Agent and Here's What Happened

CrabTrap: I Put an LLM-as-a-Judge Proxy in Front of My Production Agent and Here's What Happened I was staring at my agent logs at 10pm when I saw a response that made my stomach drop: the model had returned a code block with rm -rf wrapped in markdown. It wasn't malicious — it was a directory cleanup suggestion with just enough context to look reasonable — but my agent was one exec() away from running it without asking. That was Wednesday. Thursday I installed CrabTrap. My thesis after 72 hours