Intelligence scales. Organisations still have to exist.

The one-agent company is hiding a company inside the prompt #

Tell a single model to "run the company" and watch the instruction balloon. It needs the current goals, the facts you've agreed on, your customers, budgets, logins, unfinished promises, quality bar, and some way to tell advice apart from permission. It needs to know which work can just happen and which decision has to come back to you. It needs history, but not all of it. It needs an org chart, but only if the chart reflects who actually does what.

Calling all of that "context" doesn't make it temporary. The moment a fact has to survive swapping the model, it's company state. The moment one worker may do something another may not, it's authority. The moment two tasks can run separately but have to end up as one answer, it's coordination. The universal agent didn't get rid of the institution. It just squashed it into one giant, invisible, very hard to audit prompt.

This works great in demos, because demos end. Real companies pile up commitments. A customer still expects an answer next Tuesday even if you change model provider on Monday. A design you rejected shouldn't come back because a new session never heard about the rejection. A payment that timed out can't be safely retried just because the latest model sounds confident about it.

Abundance makes the mess bigger #

When analysis is expensive, you ration it. A small team looks at three opportunities because it can't afford to look at thirty. When analysis gets cheap, you don't just get the same three answers faster. You open more doors: more markets, more versions, more experiments, more personalised everything.

The number of things you could do grows way faster than your ability to judge them. Intelligence removes a production limit and exposes a selection limit. A system that only counts finished agent runs can actually make a company worse, because it produces more reasonable-looking work than anyone can check, reconcile or use.

So the number I care about isn't model capability on its own. It's accepted, useful output per unit of your attention, money, time and risk. Every part of that can go the wrong way while a benchmark goes up. A smarter model tries more things. A bigger team talks more. A nervous system asks for more approvals. None of it helps unless the company gets better results without dumping the new complexity on you.

• Smarter workers make it affordable to have specialists you couldn't justify before.
• More parallel work means more dependencies, contradictions and duplicated effort.
• More generated evidence means more need to tell today's proof from last week's.
• More possible actions make authority and priorities the bottleneck, not raw execution.

Departments aren't just bundles of skills #

It's tempting to think a broad enough model makes departments pointless, since it can do legal, code, sales research and marketing. But departments don't exist only because different people know different things. They hold different responsibilities, time horizons, relationships and decision rights.

Finance and product can use the exact same intelligence and still disagree, legitimately. Finance protects the runway. Product protects the users and the direction. Sales will happily promise custom work that engineering should push back on. Legal will flag a risk the owner decides to take anyway. None of these are reasoning bugs waiting for GPT-whatever to fix. They're tensions that come from what the company actually owes people.

So an agent company needs stable roles only where a recurring responsibility makes the outcome better. Restless deliberately doesn't spin up a persistent persona for every task. A temporary specialist is fine for a one-off analysis. But when a responsibility keeps coming back (owning the portfolio, looking after the company computer, guarding the money, remembering a customer), continuity becomes part of the job.

Models change. Your company shouldn't change with them #

The model market moves stupidly fast. You might switch providers for price, speed, quality, privacy or just because one is down. You might use different models for visual taste, code, sorting and quick routine work. If the company's personality is really just one model's habits, then changing a route quietly changes your culture.

Restless treats models as labour. The company's identity sits above them: the facts you've accepted, your voice, your visual language, how you behave, who's allowed to do what, and what you've decided before. These aren't prompts carved in stone. They have sources, revisions and an owner. They can change, but not because whichever session spoke last happened to rewrite them.

This also makes comparing models fairer. If two models get two different hidden versions of your company, you can't compare what they produce. A stable layer gives every worker the same picture of the company, then adds what's specific to their job. The model brings judgement. It doesn't get to redefine the company by accident.

Structure only earns its place when it closes responsibility #

There's an opposite mistake, too: assuming more structure means more mature. Agent systems love organisational theatre. Dozens of named agents, ceremonial handoffs, managers managing managers, dashboards full of things moving. Every extra boundary costs briefing, waiting, merging and reviewing.

Restless puts a boundary where accountability closes, not where a task looks big. You give an outcome to one Exec. The actual work goes to one accountable lead. The lead can bring in specialists when that genuinely pays for the extra communication: independent checking, real specialisation, or parallel work that finishes on its own. One person or agent stays responsible for bringing the result back.

The case for organisation isn't a case for bureaucracy. Stable identity, clear responsibility and work that lasts can all live alongside small teams and doing the thing directly. Structure should make complexity visible and contained, not multiply it.

My own evidence isn't all flattering #

I tested this. On several kinds of clear, bounded work, I compared a strong worker on its own against the same worker sitting behind an Exec and a lead who didn't produce anything. The quality came out the same. The supervised version took a lot longer and cost a lot more, and it didn't win on quality or recovery anywhere.

That killed a convenient version of my own argument. A lead doesn't add value just by existing. A model turn that restates a clear brief, says "looks good" to something that already passed, or relays "done" is overhead with a job title. So I kept the accountable lead but changed the default: one capable worker does the job end to end, and the lead only wakes up when something actually needs judgement. Ambiguity, changed requirements, failed checks, risky actions or a call that belongs to you.

I think the result makes the bigger argument stronger by making it narrower. The valuable thing isn't a chain of minds passing notes. It's responsibility that lasts plus intelligence used selectively. Plain software can carry routine facts. Judgement should show up where the facts stop deciding the next move.

When one agent really is enough #

A bounded task with clear inputs, one output, cheap checking and nothing lasting at stake probably needs no organisation beyond knowing who did it. Adding roles and handoffs would make it worse. Same when one capable worker can hold everything in their head and finish without running out of attention.

Structure should grow only after real work keeps showing you a boundary worth keeping. Restless doesn't create a department because a diagram says there should be one. It creates lasting responsibility when continuity, specialisation, throughput, independent checking or limiting the blast radius has actually proved useful. Otherwise the work stays direct.

So here's my practical answer to the one-agent question. Sometimes one agent is the company for the length of a task, and that's fine. What it shouldn't become is the only place where the company's truth, authority and unfinished promises live.

The scarce thing is knowing what you want #

Abundant intelligence doesn't tell a company what to want. It doesn't settle every fight between growth, trust, speed, craft and safety. It doesn't decide which promise wins when two of them collide. Those calls come from history, relationships, appetite and responsibility.

An organisation earns its existence by making those constraints useful. It gives intelligence a place to work, something that lasts to remember, a boundary to move inside, and one route for the result to come back. If all it does is add more messages, it's failed.

The company of the future might have fewer people and a lot more machine workers. It might swap those workers every week. It might move so fast that today's meetings would be impossible. None of that makes organisation disappear. It just makes it really obvious whether yours is any good.