You Can Now Yell at Claude While It's Working

The old loop went like this: you kick off a task, Claude starts doing something, you immediately see it go sideways, and you have two options — let it finish wrong and then correct it, or interrupt it, losing all the momentum and context it had built up. Neither option is good. You're either cleaning up a mess or you're back at square one, re-explaining everything from scratch.

That loop is over.

You can now steer Claude in realtime, mid-run, without blocking it. It's working, you see something you don't like, you say something, it adjusts — and keeps going. The task doesn't restart. The context doesn't evaporate. You're not waiting for it to finish a wrong thing so you can ask it to do the right thing.

This sounds like a small quality-of-life improvement. It isn't.

The blocking model forced a particular kind of human-AI collaboration — one where you had to make all your decisions upfront, like filing a complete brief before the lawyer would pick up the phone. You had to anticipate every branch the agent might take and pre-correct for it, which meant you had to think like the agent before the agent started thinking. The cognitive load was backwards. You were doing the hard work before the agent had shown you anything.

Realtime steering flips this. The agent shows you what it's doing, you react to what you actually see rather than what you imagined you'd see, and the whole thing starts to feel less like delegating and more like collaborating with someone who happens to be much faster than you at the execution part.

The analogy that keeps coming to mind is pair programming, except your pair doesn't need you to drive. They're already driving. You're in the passenger seat and you can finally grab the wheel without causing an accident.

There are failure modes here, obviously. The person who interrupts every three seconds to redirect based on vibes will produce worse results than the old blocking model — you'll get a model that's perpetually half-done, context-confused, chasing every new direction before it's finished with the last one. Realtime steering rewards people who know when to stay quiet.

But that's a skill you can develop. The old system punished you for not knowing the future.