There is a tidy story about the end of organisations. Models improve, the interfaces disappear, and one sufficiently capable agent absorbs the company. Departments become prompts. Managers become routing logic. Coordination becomes an embarrassing tax left over from a period when intelligence was scarce and unevenly distributed.
That story mistakes intelligence for organisation. Intelligence can analyse a market, repair a build, negotiate a schedule, or draft a contract. An organisation has to decide which market matters, who may accept the contract, what evidence counts, which commitments survive a staffing change, and who answers when several sensible actions conflict. Those are not merely harder inference problems. They are durable relationships between authority, memory, work, and consequence.
Restless begins from the opposite expectation: better models will increase the need for organisational structure. Cheap intelligence makes more projects plausible, more specialisation affordable, and more parallel work tempting. It lowers the cost of producing an answer much faster than it lowers the cost of deciding which answers belong together.
The universal-agent argument hides a company inside the prompt
Ask a single model to “run the company” and the instruction immediately expands. It needs current goals, accepted facts, customers, budgets, credentials, unfinished obligations, quality standards, and a way to distinguish advice from authority. It needs to know which work can proceed quietly and which decision must return to the owner. It needs history, but not every historical token. It needs the current organisation chart, but only if that chart reflects actual responsibility rather than decorative roles.
Calling all of this context does not make it temporary. The moment a fact must survive a model replacement, it is company state. The moment one actor may do something another may not, it is authority. The moment two tasks can proceed independently but must eventually produce one answer, it is coordination. The universal agent has not eliminated the institution. The institution has been compressed into an enormous, implicit, and difficult-to-audit prompt.
This compression works in demonstrations because demonstrations end. A real company accumulates commitments. A customer expects an answer next Tuesday even if the model provider changes on Monday. A rejected design should not reappear because a new session never saw the rejection. A payment that timed out cannot be retried safely merely because the latest model sounds confident. Continuity turns hidden structure into unavoidable structure.
Abundance increases the coordination surface
When analysis is expensive, organisations ration it. A small team investigates three opportunities because it cannot afford to investigate thirty. When analysis becomes cheap, the team does not simply receive the same three answers sooner. It opens more branches: additional markets, alternative implementations, narrower customer segments, more experiments, more personalised material, and more frequent review.
The number of possible actions grows faster than the owner's capacity to judge them. This is the abundance paradox. Intelligence removes a production constraint and exposes a selection constraint. A system that optimises only for completed agent runs can therefore make the company worse: it produces more locally reasonable work than the institution can reconcile, verify, or use.
The relevant unit is not model capability in isolation. It is accepted economic output per unit of owner attention, cost, time, and bounded risk. Restless uses that ratio because every term can move in the wrong direction while a benchmark score rises. A smarter model may attempt more. A larger team may communicate more. A cautious system may request more approvals. None of those changes is valuable unless the company returns better outcomes without transferring the new complexity to the owner.
- More capable workers make previously uneconomic specialisation affordable.
- More parallel work creates more dependencies, contradictions, and duplicated effort.
- More generated evidence increases the need to distinguish current proof from stale proof.
- More possible actions make authority and prioritisation—not raw execution—the limiting resources.
Departments are not bundles of skills
It is easy to imagine that a sufficiently broad model makes departments obsolete because it can perform legal analysis, write software, research prospects, and design a campaign. But departments do not exist only because different people know different things. They hold different responsibilities, time horizons, relationships, and decision rights.
Finance and product can use the same underlying intelligence and still disagree legitimately. Finance protects runway and reporting integrity. Product protects user value and coherent direction. Sales may accept a degree of custom work that engineering should resist. Legal may identify a risk that the owner knowingly accepts. These are not reasoning errors waiting for a more capable model. They are tensions created by the company's actual obligations.
An agent organisation therefore needs stable roles only where recurring responsibility improves outcomes. Restless deliberately avoids creating a persistent persona for every task. A temporary specialist may be enough for a bounded analysis. But when a responsibility recurs—owning the portfolio, maintaining the runtime, protecting financial authority, remembering a customer relationship—continuity becomes part of the work.
Changing models make stable identity more valuable
The model market changes quickly. A company may switch providers for cost, latency, capability, privacy, or availability. It may use different models for visual judgement, code production, classification, and fast routine work. Treating one model's behavioural tendencies as the company's identity creates a fragile organisation: a routing change becomes an unannounced culture change.
Restless treats models as labour. The company identity sits above them: accepted product truth, voice, visual language, behavioural norms, authority, and precedent. These records are not prompts carved in stone. They have provenance, revisions, and an owner. They can evolve without being silently rewritten by whichever session happened to speak last.
This separation also makes model comparison more honest. If two models receive different hidden interpretations of the company, their output cannot be compared cleanly. A stable institutional layer gives each worker a compact common operating picture, then adds focused context for its responsibility. The model is allowed to bring judgement. It is not allowed to redefine the company accidentally.
Organisation is valuable only when it closes responsibility
There is an opposite mistake: assuming that more structure is automatically more mature. Agent systems can manufacture organisations as theatre—dozens of named workers, ceremonial handoffs, recursive managers, and dashboards full of motion. Every additional boundary creates briefing, waiting, integration, and review costs.
Restless places an organisational boundary where accountability closes, not where a task merely looks large. The owner gives an outcome to one Exec. Productive work is delegated to one accountable lead. The lead may commission specialists when independent evidence, genuine specialisation, or locally closing parallel work can repay the communication cost. One actor remains responsible for returning the result.
This distinction matters because the case for organisation is not the case for bureaucracy. Stable identity, explicit responsibility, and durable work can coexist with sparse teams and direct execution. The institution should make complexity legible and contain it—not multiply it.
The evidence is not uniformly flattering
Restless's own EXP-17 compared a strong Codex worker operating solo with the same worker behind an Exec and a non-producing lead. Across four valid sparse pairs, outcome quality reached parity. The supervised chain cost 1.73 to 3.30 times as much and took 1.77 to 2.94 times as long. It produced no quality or recovery win in those tasks.
That result rejects a convenient version of the thesis. Organisation does not add value merely because a lead exists. A model turn that restates a clear brief, acknowledges a passing artifact, or mechanically relays a completion fact is overhead wearing a job title. Restless kept accountable leadership but changed the operating prior: one capable end-to-end producer is the default beneath a lead, and supervisory intelligence wakes for material ambiguity, changed requirements, failed evidence, unsafe effects, or owner-facing judgement.
The experiment strengthens the larger argument by narrowing it. The valuable institution is not a serial chain of minds. It is durable responsibility plus selective intelligence. Deterministic machinery can carry routine facts. Organisational judgement should appear where the facts no longer decide the next move.
When the single agent really is enough
A bounded task with clear inputs, one artifact, cheap verification, and no durable consequence may need no organisational machinery beyond attribution. Adding roles and handoffs would reduce performance. The same is true when a single capable worker can hold the relevant context and close the outcome without exhausting its attention.
The architecture should grow only after repeated work reveals a boundary worth preserving. Restless does not create a department because a taxonomy says one should exist. It creates persistent responsibility when continuity, specialisation, throughput, independent evidence, or blast-radius control has demonstrated value. Otherwise the work stays direct.
This is the practical answer to the universal-agent question. Sometimes one agent is the company for the duration of a task. What it should not become is the only place where the company's truth, authority, and unfinished obligations exist.
The scarce thing is coherent direction
Intelligence becoming abundant does not tell a company what to want. It does not resolve every conflict between growth, trust, speed, craft, and safety. It does not decide which promise should survive when two promises collide. Those decisions draw on history, relationships, appetite, and responsibility.
An institution earns its existence by making those constraints productive. It gives intelligence somewhere to work, something durable to remember, a boundary within which to move, and one path by which the result returns. If it merely adds more messages, it has failed.
The future company may employ fewer people and many more machine workers. It may change those workers weekly. It may execute at a speed that makes today's coordination rituals impossible. None of that makes organisation disappear. It makes the quality of organisation visible.