← Home

YOLO-mode by default

I now only run agents in full-permission mode. I do this all the time, despite the UI warnings, despite me knowing that this is a generally bad idea, and despite there being a known golden path, sandboxing.

7 months ago, around January, when I first started heavily using agents at work, I would never have done this.

I think it’s worth reflecting on this behavioral change as a way to reason about the past 7 months and about larger trends this points to. This is anecdotal, but I’m a strong believer in: what is personal is universal.

I run agents in full-permission mode for three reasons.

  1. I can get work done faster.
  2. I have more trust in the models.
  3. I haven’t bothered to run remote agents or figure out more secure sandboxing.

I get work done faster in YOLO mode because agents are not blocked on permission gates. Permission gates effectively serialize work. They rely on my approval and I can only work through gates one-at-a-time, sequentially.

I have more trust in the models these days. Over time, models and harnesses have improved. I can prompt with less care and with larger scope and still have decent, functional results. Of course, it’s not perfect, the underlying system is often a mess if not tended to, but it’s good enough for many early projects and prototypes.

I haven’t bothered to sandbox or run cloud agents. I should. I don’t know why I haven’t done this other than I haven’t felt the urgent need. I’ll try to do this more often. It’ll be more secure and, if in the cloud, will be easier to run background tasks.

Now as for future trends and predictions.

I think there’s a ticking time bomb aspect to YOLO mode. I think many people are running in YOLO mode, and I get the inkling we’ll see a pretty terrible exploit within the next year. I think there’s appetite for a default, well-understood sandboxing effort. I think the sandboxing solution here will be a free, open standard, not something that can be monetized. I think cloud agents will see good adoption but, depending on their economics, may not be a default running-mode for agents.