Human-in-the-loop runs out of human.
Approving every step doesn't keep you in control at scale. It burns through the attention you need for the calls that actually matter.
"Keep a human in the loop" sounds like the responsible thing to say. The agent does stuff, you check the stuff, nothing bad happens. Lovely.
Then you actually try it. Approve this file read. Approve this command. Approve this webpage. By prompt forty you're clicking yes the way you accept cookie banners, and congratulations, the human in the loop is now a very slow rubber stamp. Every approval costs reading time, every notification breaks your focus, and every vague request makes you rebuild context the system already had. As the machines get faster, you either become the bottleneck or you stop thinking before you click. Both defeat the point.
Restless is built around a different shape: lots of freedom for the machines on the inside, human authority at the edges. That's not a slogan about autonomy. It's a design for protecting the one thing human oversight actually runs on, which is attention.
Lots of approvals isn't the same as control
A system can ask permission for every tool call and still leave you with no idea what's going on. You see a stream of tiny actions: read a file, run a command, fetch a page, edit a draft. Each one makes sense on its own, but what they add up to doesn't. You're approving motion without understanding direction.
Constant approvals also train you. When ninety-nine prompts are routine, the hundredth gets the same reflex. Security people call this warning fatigue. Managers call it rubber-stamping. The software can honestly record that a human approved it while the human's contribution has become pure ceremony.
The thing to measure isn't how often a person shows up in the workflow. It's whether the system brings them the decisions that need their judgement, with enough evidence to decide well. A hundred "yep, fine"s can contain less real control than one well-prepared question.
Your attention is part of the budget
Engineering budgets usually track compute, tokens, speed and money. Your attention gets left out because it's hard to measure. But it's often the tightest limit there is. You can buy more inference. You can't buy a second uninterrupted morning.
Attention costs you in at least three ways. There's the interruption: dropping what you're doing and loading something new. There's the catch-up: working out what happened before the question arrived. And there's the judgement itself: weighing options when you're not sure. Bad agent systems hand you all three and then report that they saved you time.
So a useful attention budget tracks more than message counts. It asks whether the interruption was necessary, whether the decision showed up ready, whether similar questions could have been batched, and whether the company learned enough not to ask again.
Work out what needs you from facts, not vibes
Lots of systems decide what's urgent by asking a model to score a message. That's fine for sorting a list, but it's a shaky foundation for authority. A confident model might call its own favourite action low-risk. A nervous one labels everything important. Your queue just becomes a reflection of the model's personality.
Restless starts from facts owned by whatever is actually responsible for them. Is a needed login missing? Did an automatic check reject the output? Does the next step go beyond what's allowed or budgeted? Is something out in the world in an unknown state? Did you explicitly say you want to make this call yourself? Did the evidence turn up a real trade-off that no existing rule settles?
Those facts don't replace judgement. They decide whether judgement is needed and where the answer has to be written back. A permission request gets answered where permissions live. A handoff to you gets answered against the work it came from. Your cockpit shows both in one place without becoming a second owner of either.
- Routine work that's within the rules just keeps going, without asking for applause.
- A missing permission becomes a specific request, not a vague plea for help.
- An unknown outcome out in the world gets checked before anything is retried.
- A strategic question comes back with options, evidence and what waiting will cost.
Decisions should show up ready
Delegation fails when the hard bit lands on you raw. "Which option do you prefer?" is a pretty useless question if the system could have compared the options, tested the assumptions and worked out what each one leads to. The machine saved itself effort by passing the uncertainty up to you.
A prepared decision should include what it suggests doing, the strongest evidence for it, a credible alternative, what's still uncertain, what permission it needs, whether it can be undone, and what happens if you wait. It should link to the actual thing you're judging (the document, the build, the page). The logs are still there if you want them. They're just not what you see first.
This changes your job. You're not babysitting a process minute by minute. You're bringing taste, responsibility, relationships and appetite for risk to the exact point where those human things change the result.
Taking part isn't the same as having authority
A person can take part in a task without having authority over it. They might answer a factual question, give a taste opinion or show how something's done. And a machine can prepare an action perfectly while still not being allowed to do it. Blur the two and you get either pointless interruptions or unsafe autonomy.
Restless treats authority as its own boundary. Models can explore, edit, test, browse and prepare inside the environment they've been given. Identity checks, extra permissions, budget limits, consequential actions and the official record of them all stay in plain, predictable software. The model can explain why it's asking. It can't approve its own promotion.
That's lopsided on purpose. The machine gets a lot of freedom over work that can be undone, and very little over consequences. You don't need to approve its path through every file edit. You keep the decisions that change the company's promises or the outside world.
Quiet systems still need a big red button
Fewer notifications can't mean hiding the company from you. You still need to be able to see who's responsible for what, freeze anything with consequences, stop an agent, take away a permission, jump into the company computer or roll back internal state. Rescue controls matter precisely because they're not the normal way of working.
The main screen should stay calm: what's happening, the real decisions, the real risks and what's next. Deeper views can show roles, prompts, spend, messages, permissions and logs when you need to dig in. Hiding detail until you ask isn't about looking minimal. It's about protecting your ability to see the few things that matter right now.
Silence needs to be explained, too. "Nothing's happening because everything's fine" should look different from "nothing's happening because a wake-up got lost". Cutting down on noise can't make a broken system look exactly like a peaceful one.
Some work needs a human in the room the whole time
Not every workflow should push the human to the edge. Negotiation, therapy, hiring, sensitive leadership and exploratory creative work can depend on an ongoing relationship and interpretation. The human isn't a final gate bolted onto mechanical production. They're part of what the work is.
The mistake would be turning "humans at the edge" into a rule for everything. Restless should work out where the judgement actually lives. If the human interaction is where the value comes from, the system can prep context, capture what was agreed and cut admin while keeping the person in the middle of it.
So the attention argument is about removing coordination you didn't need, not about minimising humans for the sake of it. Good automation makes your attention more deliberate. It doesn't treat your attention as waste.
Keeping AI on track needs a human who's awake
A lot of talk about AI alignment is about whether the system follows instructions. In a company, instructions are always incomplete and sometimes contradict each other. The more useful question is whether the risky, unresolved stuff reaches someone who can judge it, before their attention has been used up by noise.
A human who's technically in every loop but mentally absent from most decisions isn't a safety system. A system that does routine work quietly, keeps the evidence and stops at a clear permission boundary can be more autonomous and more controllable at the same time.
The point of Restless isn't to remove you. It's to stop spending you on work machines can already do, so you're still around for the decisions that make the company yours.