How much can you delegate to agents?

The article tried to bring order to the chaos with a simple rule: don’t trust an AI helper more just because the software got smarter. Instead, trust it based on the job. If it’s easy to check and easy to undo, let it run. If it’s hard to judge or expensive to fix, keep a human close. Sensible? Yes. But the comments quickly turned this from a calm guide into a full-blown trust issues convention.

The loudest reaction was basically: “You’re missing the real danger.” jolaflow argued that checking whether the work “passes” isn’t enough if the AI has quietly drifted away from the original goal. In other words: sure, it finished the task, but did it finish the right task? That hit a nerve. Others piled on with a more social fear: as apical_dendrite put it, what if people are just “pissed” because they expected thoughtful work and got AI slop instead? Ouch.

Then came the practical operators. ChicagoDave described a very relatable survival strategy: sometimes the bot works in a safe sandbox, sometimes it’s direct on the machine, and sometimes you keep “holding the steering wheel” because you just know it might wander off. That image basically became the thread’s unofficial meme. And right in the middle of it all, ChulioZ dropped the debate grenade: what’s the real difference between a human review and a human safety gate? Suddenly the neat four-level system looked less like a map and more like the start of a comment war.