Remote Control Is a Force Multiplier, Not a Convenience

Remote control is the most important change to how I use AI this year, and it's disguised as the least important.

On paper it's a convenience: start a task on your machine, check on it from your phone. A lot of people - me included, at first - saw exactly that and shrugged. That read is wrong, and the distance between how minor it looks and how much it actually changes is the whole point. It matters for a reason that isn't on the feature list.

The point isn't mobility. It's that the agent keeps working when I don't.

Here's what actually changed. I run Claude Code on Claude Max - the 20× tier, $200 a month, about $250 once Norwegian VAT is on top. And the thing about an agent you can reach from anywhere is that it can be working anywhere - which really means it can be working any-when. While I sleep. While I stand in a queue. On a flight. At lunch. At 3am when I can't sleep and an idea won't leave me alone.

Remote control is how I point that idle time at a problem. Someone on Bluesky put the everyday version of it well: "leaving the office while it's still cooking, checking status from mobile." That's all it is, really. The work doesn't stop because I've walked away.

So I leave things cooking. Overnight, I'll set a long experiment running - self-improving code that just keeps trying until it has exhausted every way I can think of to improve some specific thing. I'll point it at a big messy dataset and have it cross-reference until something interesting falls out. I'll have it write the small utility I need for work, or for the house, that I'd never have found an afternoon for. Some runs are just demos to prove to myself that a thing is possible now. And I'll run models against each other - adversarial bake-offs to see which one actually does what - and feed the results back into my own system prompts, which I'm tuning more or less continuously.

About the bill

I usually max out my weekly allowance. Lately it's felt like I have more room, and the honest question is why. The tempting answer is that Anthropic quietly raised the limits. I went looking, and I can't find any sign that they did. What I found instead was the opposite: most people badly underuse what they're paying for. One developer measured himself at "~22% of my weekly cap on average - paying for 5×, behaving like 1×" and built a little tool to shame himself into using it.

Paying for 5×, behaving like 1×. That's the normal case. Remote control is how you stop being it.

So the room I feel isn't a bigger cap. It's that remote control plus sensible model routing means the agent is working during hours I'm not - and an agent that runs while you sleep saturates a plan in a way an agent that only runs while you watch never will. I'm not getting more allowance. I'm finally using the one I have. (If you want the number, Claude exposes a local usage endpoint you can read - it's a sobering thing to actually look at.)

There's a piece of advice here I know I should follow more faithfully than I do. Boris Cherny, who created Claude Code, makes the case for not over-engineering your prompts at all: the more general model beats your specific scaffolding, so you build for the model six months out and let the new one work on its own. He's right, and every model jump proves it again. But I draw one line. The context that's specific to my work and my projects - what I've learned about my domain, my codebase, the constraints nobody else hits - is hard-won, and it doesn't arrive in a model update. The model gets smarter on Anthropic's schedule; my understanding of my own problems only gets better because I keep it. So I throw out the scaffolding and keep the knowledge.

Why I'm writing any of this down

Here's the part that's harder to say. I can see where this is going - the change, and what it means - clearly enough that it's frustrating to watch how slowly it lands for everyone else. I catch the signs that people are, little by little, coming around. But it is slow. Maddeningly slow. And there isn't really an outlet for the thing I see in my head when I'm standing in an airport queue with an agent running in my pocket.

So this is the outlet. All I can actually do is keep learning, keep practising, keep using this stuff - and write down what I see while it still looks like nothing much to most people. It is, as I keep saying in these notes, all about agentic: not the model's cleverness, but what happens when you let it act, at length, on your behalf.

Which, lately, is mostly in queues and at 3am. I've made my peace with that.