The Real Bottleneck Was Never Time. It Was Attention.

There is a moment in every startup when the work stops feeling like work and starts feeling like atmospheric pressure.

Not one dramatic collapse. Not one clean failure. Just accumulation.

Too many tabs. Too many threads. Too many half-decisions. A backlog that looks organized but behaves like weather. One person is fixing a deployment issue. Another is untangling an edge case. Someone else is translating a client request that arrived already broken into calls, chats, tickets, assumptions, and urgency. Founders stop being founders and become routers. Senior people stop being specialists and become context-recovery engines. The company grows, but its internal logic does not grow with it.

For a while, this feels like speed.

Then it starts to feel like heat.

And eventually, if nothing changes, it becomes noise.

That is why I keep returning to the same set of ideas: task prioritization not as a ranking exercise, but as a discipline of attention allocation; attention purity as a measurable and protectable asset; and the need for a generalized or specialized Company OS that can hold all of this together.

The calendar is not broken. The coordination layer is.#

In startup and scale-up environments, people love to say that time is scarce. That is true, but incomplete.

Time is not the first resource that breaks.

Attention is.

Because time can be scheduled. Attention can be contaminated.

A four-hour block on the calendar means almost nothing if it is broken by notifications, side quests, missing context, unresolved ambiguity, and the low-grade social pressure to remain constantly available. A full day can disappear into responsiveness without producing a single clean unit of thought. Microsoft’s 2025 workplace telemetry described heavy users being interrupted every two minutes during core work hours, which is a good way to explain why so many teams look busy while remaining cognitively underpowered.¹

This is where most prioritization systems quietly fail. They assume that the unit being allocated is time. It usually is not. The unit being allocated is clean cognition under constraint.

The startup pain follows from that. You are not just choosing what to do next. You are choosing what deserves scarce, high-quality attention in an environment that keeps fragmenting it. That is a coordination problem before it becomes a planning problem.

Remote work did not break the system. It exposed it.#

I have worked remotely for years, and the main problem was almost never whether work got done. The harder part was building connection deliberately instead of inheriting it accidentally from proximity. That difference matters.

For many people, remote work is not just a perk. It is a way of protecting the conditions under which serious work becomes possible. It removes commuting drag. It lowers performative busyness. It gives people more control over the shape of the day. It can reduce interruption density and let work happen closer to actual cognitive peaks instead of office theater. Gallup’s tracking shows that among remote-capable workers, most want hybrid work, about one-third prefer fully remote work, and fewer than one in ten want to be fully on-site.²

The old managerial reflex says: if I cannot see you, I cannot trust that you are working.

The data are less nostalgic than that reflex.

In a randomized field experiment at Trip.com, hybrid work improved job satisfaction, cut quit rates by one-third, and did not reduce performance or promotion rates. Just as interestingly, managers started the experiment expecting hybrid work to lower productivity by 2.6% on average and ended it believing it improved productivity by 1.0%.³ Earlier work on the company’s call-center operations found a 13% productivity increase under work-from-home conditions, driven by both more minutes worked and higher productivity per minute in a quieter environment.⁴

At the same time, not all remote settings work for all work or all people. Research on collaboration patterns found that firm-wide remote work can make networks more static and siloed, with fewer bridges across groups.⁵ A broader Stanford-SIEPR summary argues that the productivity effect depends heavily on the mode: hybrid appears roughly neutral on productivity in many knowledge-work settings, while fully remote arrangements are often associated with lower productivity on average because of frictions around communication, mentoring, and innovation.⁶ Personnel data from a large IT-services firm also found that when work moved home during the pandemic, hours rose, uninterrupted work shrank, coordination time increased, and measured productivity fell between 8% and 19%.⁷

So the honest conclusion is not that remote work is always better, or always worse. It is that work mode interacts with task structure, team design, trust, seniority, and operating discipline. Remote work did not create most coordination problems. It made them impossible to hide.

That is also why remote-work debates so often go wrong. They are presented as ideological fights when they are usually system-design fights. The real question is not whether people are visible. It is whether the organization is designed so that people can do focused work, stay connected, and make good decisions without drowning in interruption or isolation.

GitLab is a useful case here, not because every company should copy it, but because it shows what it looks like when a remote organization makes its ways of working explicit. Its public handbook documents the mechanics of operating as an all-remote company at significant scale instead of relying on ambient office folklore.⁸

Small teams matter because trust scales poorly by accident#

This is one reason I think small teams matter more than many companies admit.

Smaller teams reduce coordination overhead. They make context easier to preserve. They let people know one another beyond ticket boundaries. They make it easier to notice who is overloaded before the burndown chart notices. They also make it more likely that work and life can meet at a human scale: not as invasive oversharing, but as enough mutual context that care becomes operational rather than ceremonial.

That matters. People do not arrive at work as clean abstractions. They arrive carrying sleep debt, finances, family obligations, illness, grief, medication, hope, irritation, and whatever happened fifteen minutes before the first meeting.

Software-engineering research has repeatedly found that larger teams face stronger coordination costs and lower productivity per person as team size grows.⁹ Research on virtual teams likewise keeps returning to the same ingredients: trust, knowledge sharing, and psychological safety are not cultural extras but performance variables.¹⁰¹¹¹²

A good small team does something the org chart rarely captures: it lets people matter to one another in ways that improve both work and daily life.

Attention purity#

This is where I use the phrase attention purity.

Attention purity is the degree to which a person or team can engage with a problem without contamination from irrelevant interrupts, unresolved ambiguity, missing context, unnecessary coordination noise, or constant role-switching.

It is not merely focus.

It is cleaner than that.

It means the objective is legible. The context is accessible. The ownership is known. The dependencies are visible. The time horizon is protected. The person doing the work is not being asked to rebuild the same mental model every twenty minutes.

Not all work degrades equally under fragmentation. Some work survives interruption. Other work rots inside it.

Architecture rots inside it. Strategy rots inside it. Research rots inside it. Careful prioritization rots inside it. Writing rots inside it. Sensitive conversations rot inside it.

Once low-purity attention becomes the norm, organizations start optimizing around whatever survives in low-purity conditions: reactive work, cosmetic metrics, constant checking, over-coordination, and decision-making by volume instead of clarity. The tragedy is that this often gets mistaken for professionalism.

Priority is never singular#

A startup backlog is an MCDA problem wearing a Kanban costume.

What many teams call prioritization is usually a simplified ranking ritual pretending the problem has one dimension. But real operating decisions almost never do.

They live inside tension: customer value versus implementation cost, speed versus reversibility, short-term revenue versus long-term positioning, technical debt versus opportunity capture, learning value versus certainty, morale versus pressure, local optimization versus system coherence, and increasingly, throughput versus attention fragmentation.

That is exactly the terrain that multi-criteria decision analysis was built for: multiple alternatives, conflicting objectives, mixed quantitative and qualitative evidence, and stakeholders who do not value the same criteria in the same way.¹³¹⁴

Seen this way, attention purity is not just a poetic phrase. It becomes one of the criteria.

A task may deserve priority not only because it is urgent, but because it is compounding, emotionally expensive, operationally dangerous, difficult to reverse, or dependent on a rare kind of uninterrupted cognition that the organization keeps failing to protect.

So the question stops being, “What should we do next?”

The better question becomes, “Given what matters, what it costs, what it risks, what it blocks, what it teaches, and what kind of attention it requires, what deserves clean execution now?”

Mental health is not adjacent to decision-making. It is inside it.#

Mental health is still too often treated as a side topic, something to be delegated to HR after the “real” decisions have already been made.

That is a category error.

People are not compute nodes. They are embodied decision systems.

Mental health changes judgment quality. It affects patience, memory, communication, frustration tolerance, interpretation of ambiguity, and the ability to stay with a hard problem without collapsing into panic, defensiveness, or avoidance. The World Health Organization is explicit that poor working conditions — including excessive workload, low control, and discrimination — can harm mental health, and it estimates that depression and anxiety lead to 12 billion lost working days every year globally.¹⁵ A 2024 systematic review likewise found detrimental effects of occupational stress, shift work, and probably prolonged working hours on cognitive functioning.¹⁶

Remote work fits here too. Flexibility can improve autonomy and make life more livable. But remote work can also blur boundaries and intensify loneliness when the social layer is neglected. Gallup’s 2025 analysis found that fully remote workers reported higher engagement than some on-site groups, but they were less likely to be thriving overall than hybrid workers and more likely to report anger, sadness, loneliness, and stress.¹⁷

This is why a serious operating model cannot treat emotional reality as noise. Daily life is part of the decision environment.

At the same time, none of this justifies workplace surveillance. The EU AI Act prohibits emotion-recognition systems in workplaces and education institutions, except in narrow medical or safety contexts, and requires effective human oversight for high-risk AI systems.¹⁸¹⁹ The answer is not to measure workers more invasively. The answer is to stop designing systems that predictably grind them into worse decisions.

Where automation fits — and where it should not#

Some kinds of work are structurally friendly to automation.

The pattern is surprisingly consistent: standardized inputs, repeatable loops, self-serve or near-self-serve onboarding, clear success criteria, cheap or reversible errors, and exceptions that are rare enough to classify instead of negotiate from scratch every time.

Other work resists that structure.

Top-level strategy under uncertainty. Major negotiations. Final legal, privacy, or compliance judgment. Severe incident command. Hiring, firing, promotion, and compensation. Crisis communication. Rights-affecting decisions. Decisions where accountable human oversight must be provable.

That distinction matters more than the hype cycle does.

The stronger near-term model is not full autonomy in the abstract. It is governed partial autonomy: automate repeatable throughput, preserve humans for consequential judgment.

Current evidence supports that caution. McKinsey’s 2025 global survey found that while experimentation with AI agents is broad, only 23% of respondents reported scaling an agentic AI system somewhere in the enterprise, and no single business function crossed 10% reporting scaled use.²⁰ Research on long-horizon coding tasks also shows a large gap between narrow benchmark success and sustained software evolution in realistic settings.²¹ Empirical studies of agent-authored pull requests suggest that these contributions are still accepted less often and reveal failure modes linked to trust, context alignment, and reviewability.²²

So yes, autonomy matters. But the useful frontier is not “the AI runs the company.” It is “the system handles more recurring throughput without pretending that accountability has disappeared.”

What a Company OS is#

Once you frame the problem this way, the missing piece becomes easier to name.

A company does not mainly lack one more tool. It lacks a coordinating layer.

That is what I mean by Company OS.

A Company OS is the continuously running socio-technical control layer of an organization: the layer that turns goals into governed execution by allocating attention, routing work and decisions, coordinating people, software, and agents, preserving institutional memory, enforcing constraints and approvals, modeling the company’s current state, and adapting operations as conditions change.^23242526

That definition is intentionally stricter than “our way of working” and wider than “our tool stack.”

A Company OS is not just a project-management stack. Not just an ERP. Not just a wiki. Not just a process diagram. Not just a digital twin. Not just an AI copilot.

It sits above and through those things.

It is closer to the nervous system.

Older literature gets part of the way there. Enterprise-engineering work has described an Enterprise Operating System as a viability-supporting, self-governing core of the enterprise.²³²⁴ Other work has approached an organizational operating system more formally by mapping basic operations, resources, and structures.²⁵ Gartner’s digital twin of an organization concept gets close from the modeling side by describing a dynamic software model that helps explain how an organization operationalizes its business model, responds to changes, deploys resources, and delivers value.²⁶

But the AI-native version needs one more step.

In the way I imagine it, a Company OS has five inseparable roles.

It is the decision layer, because it frames tradeoffs and routes judgment.
It is the attention layer, because it decides what deserves clean execution and what can wait.
It is the memory layer, because it turns past work into reusable knowledge instead of letting it decay into folklore.
It is the orchestration layer, because it coordinates humans, tools, workflows, and agents across time.
And it is the governance layer, because autonomy without policy rails, approval logic, and accountable human oversight is not an operating system but a liability.²⁷²⁸

If a company has CRMs, ERPs, chats, docs, dashboards, and agents but still depends on memory trapped in people, constant context reconstruction, noisy prioritization, and heroic intervention to keep moving, then it has software but not yet a Company OS.

One fracture, many symptoms#

This is the missing bridge in a lot of conversations: remote work, small-team trust, attention purity, MCDA, mental health, workflow automation, governance, and the Company OS are not separate topics.

They are different faces of the same coordination problem.

Remote work raises the visibility of coordination flaws. Small teams help absorb them through trust and human familiarity. Attention purity explains why some work collapses under interruption while other work survives. MCDA explains why prioritization is inherently multi-objective and cannot be reduced to a sorted list. Mental health explains why the decision-maker is never a detached machine but a living human under conditions. Workflow automation explains which recurring loops can safely be delegated. Governance explains where delegation must stop. And the Company OS is the missing layer that ties those pieces into one operational logic.

Without that layer, companies keep treating symptoms separately. They debate remote work as culture. They debate burnout as wellness. They debate prioritization as backlog hygiene. They debate memory as documentation. They debate autonomy as tooling. They debate governance as legal overhead.

But these are not separate malfunctions.

They are fragments of one broken operating system.

Memory is the hidden layer#

Most companies do not really have institutional memory.

They have storage.

History is preserved, but incoherently.

The real leap happens when memory stops being treated as archived documentation and starts being treated as operational infrastructure. That means recovering knowledge from the traces work already leaves behind: chats, tickets, docs, approvals, incidents, outputs, logs, decisions. It means turning those traces into reusable playbooks, routing rules, escalation trees, scoring rubrics, and skills.

This is one reason I do not think the important breakthrough is “bigger context windows” alone. The important breakthrough is structured memory: distilling experience into retrievable, reusable abstractions, closer to how people form procedures than how databases store logs.

The literature is still early, but it is moving in this direction. A 2025 ACL paper showed that dialog workflows can be extracted from historical interactions and reported measurable accuracy gains over baseline extraction methods.²⁹ A 2025 systematic review on unstructured data in process mining found only 24 primary studies out of 1,379 search results, which says two things at once: the field is young, and the opportunity is real.³⁰ In agent research, memory, reflection, and planning are no longer side features; they have become architectural concerns.³¹ More recent work such as Mem0, WISE-Flow, and Trace2Skill pushes the same idea further: extract, consolidate, retrieve, and evolve procedural knowledge instead of pretending every task starts from zero.³²³³³⁴

That matters because a future Company OS is not just a workflow engine.

It is a memory system with consequences.

Generalized core, specialized grammars#

I do not think one universal system will run every company in the same way.

What I imagine is a generalized core with specialized grammars.

The generalized layer handles universal organizational problems: prioritization, routing, memory, ownership, approval, escalation, simulation, auditability, attention allocation.

The specialized layer adapts those mechanisms to the domain. Different environments have different exception rates, different risk models, different evidence structures, different tolerance for automation, and different definitions of harm.

It also adapts to company stage. One useful public taxonomy here comes from IT Archetypes, which argues that organizations move through materially different operating conditions as they go from zero-to-one exploration, to one-to-ten stabilization and scaling, to ten-to-hundred replication and maintenance.³⁵ The labels used there for talent mixes — Commandos, Infantry, Police — are deliberately rough and sometimes polemical,³⁶ but the deeper insight is sound: the operating layer a company needs changes with the shape of uncertainty. Early companies need high-bandwidth generalists and rapid recombination. Scaling companies need stronger coordination, disciplined interfaces, and gradual systemization. Mature companies need repeatability, resilience, and defenses against bureaucratic drift.

So the point is not a universal machine that flattens all industries into one abstract dashboard. The point is a core operating logic that can be specialized where reality demands it — by domain, by risk class, and by stage of organizational evolution.

That is also why I keep circling around Company OS and Decision OS almost interchangeably. The first emphasizes the whole organization. The second emphasizes the judgment layer. In practice, they may become two views of the same thing.

Human oversight is not a temporary inconvenience#

A shallow version of the future imagines AI replacing managers, operators, and specialists wholesale.

I do not find that convincing.

A better version is one where the system automates throughput, surfaces tradeoffs, proposes actions, remembers patterns, routes work, and keeps operating under supervision, while humans remain responsible for strategic, rights-affecting, ethically loaded, politically sensitive, or hard-to-reverse decisions.

That is aligned with both technical reality and governance direction. NIST’s AI RMF explicitly emphasizes defining human roles and responsibilities when AI systems are used in operational settings.²⁷ The OECD’s updated AI Principles continue to center trustworthy, human-rights-respecting, human-centric AI.²⁸ The EU AI Act hardens that direction through prohibited practices and specific human-oversight requirements for high-risk systems.¹⁸¹⁹

So no, human oversight is not the embarrassing part of the design that disappears once the models get better.

It is part of the design.

Ledgers, trust, and the edges of coordination#

Blockchain is not the heart of this thesis.

But it does belong on the map.

Not as decoration. Not as ideology. As infrastructure for a certain class of coordination problems: shared state across organizations, machine-readable permissions, auditable commitments, settlement, programmable governance, and cases where trust needs to be distributed instead of assumed.

Recent enterprise research on blockchain interoperability argues that conventional interoperability approaches are not enough for blockchain-enabled business processes and that blockchain-specific frameworks are needed to handle systems, data, compliance, and cross-organizational coordination.³⁷

That is exactly why blockchain remains relevant at the edge of a future Company OS. Not because every company needs tokens. Because some organizations will eventually want stronger machine-readable governance and cross-organization auditability than conventional enterprise stacks make comfortable.

Beyond screens, still inside the loop#

The same goes for human-computer interfaces.

They are not tomorrow morning’s default enterprise tool. They are still early, uneven, and largely clinical.

But they are no longer pure science fiction either. Nature’s 2023 speech-neuroprosthesis work showed unconstrained sentence decoding from neural activity at high rates.³⁸ Nature’s 2025 reporting on near-real-time thought-to-speech systems pushed that direction even closer to conversational use.³⁹

That does not mean mind-controlled companies arrive next quarter.

It means that a longer-horizon future — one in which intention, memory capture, planning, and interaction become more fluid than today’s screen-and-keyboard model — is now reasonable to place on the horizon.

And the point, for me, is not the novelty of the interface. The point is what it connects to. A Company OS gives those interfaces something meaningful to control.

The world I imagine#

By 2030#

By 2030, the first serious versions of a Company OS will not look like fully autonomous corporations.

They will look like decision-support and memory layers sitting on top of messy organizations.

They will observe work, structure it, recover forgotten context, surface tradeoffs, protect attention, recommend actions, and help teams stop rediscovering the same lessons every month. They will be strongest in recurring workflows, narrow operational loops, and settings where success criteria are clear enough to rank, simulate, or explain.

For software businesses, the first credible version of this future is an AI-native software house that can turn conversations into plans, plans into scoped work, and scoped work into reviewed releases faster than today’s teams can manage context switching — with humans still reserved for the consequential calls.

The company will still be human-run.

But it will no longer be as human-memory-limited.

By 2040#

By 2040, I expect companies to become more legible to themselves.

Not just through dashboards, but through machine-readable governance, persistent procedural memory, workflow-aware agents, and organization-level simulation. At that point the Company OS stops looking like software sitting on top of the business and starts looking like the business’s nervous system: routing signals, preserving memory, classifying exceptions, tracking state, and coordinating humans, software, and physical operations across time and distance.

This is also where the company starts feeling less like a filing cabinet and more like a living system. Not alive in the mystical sense. Alive in the operational sense: it can observe itself, update itself, patch itself, preserve its own history, pursue work inside guardrails, and keep functioning for a while without constant managerial handholding.

You do not just open a dashboard. You consult a continuously running model of the organization and its options.

By 2050#

By 2050, the interface may stop looking like an interface.

The company may have an avatar, a voice, a visible state, simulations, forecasts, permissioned action layers, and a memory you can actually work with. You may still use dashboards, maps, sliders, boxes, scenario models, and alert thresholds. But you may also sit in your living room and talk to the system as if you were talking to the operational mind of the business.

You may ask it to explain why a roadmap shifted. You may ask it to simulate a hiring freeze, a market shock, a vendor failure, or a product expansion. You may let it run in a constrained autopilot mode for a week while you are away, with clear approval layers, escalation rules, and alerts if the situation moves outside the guardrails. It may propose options before it proposes action. It may know when to ask and when to execute.

And the scope may spill beyond the company itself. The same operational interface that helps coordinate a business may also broker services around it: scheduling, procurement, professional contact, financial coordination, personal admin, maybe one day much more. Not because every service becomes one giant machine, but because the operating layer becomes good enough at permissions, memory, negotiation, and handoff to mediate more of daily life.

That is also where blockchain and more direct interfaces become less decorative and more intelligible. Shared ledgers are one way of managing permissions, commitments, and value exchange across boundaries. Neural or other advanced interfaces are one way of collapsing friction between intention and control. Neither is the point by itself. They matter because they connect to a deeper need: an operating layer that can hold memory, decision-making, coordination, and governed autonomy together.

So yes, I can imagine a future where you are in an electric car in August, half on holiday, half still inside the loop, asking the company what changed, why it changed, and what options it recommends before anything consequential happens. I can imagine a future where ideas spoken from a sofa are captured, structured, challenged, and turned into product candidates that the organization can evaluate and build. I can imagine a future where the company can expand products, spin up new operational branches, coordinate people and agents, and still know when a human must remain the final authority.

Not because I think “AI boss” is the goal.

Because I think governed, legible, memory-rich organizational intelligence is.

In the end, it comes down to this#

The modern company is a decision system trapped inside tools that do not hold its full complexity.

It stores tasks but forgets reasons.
It measures output but ignores cognitive conditions.
It debates remote work as ideology instead of asking what conditions produce the best thinking.
It says people matter while treating their emotional reality as noise.
It wants autonomy without governance.
It wants intelligence without memory.
It wants scale without coherence.

A true Company OS — or, if one prefers the narrower phrase, a Decision OS — would begin to repair that.

It would treat prioritization as MCDA under uncertainty.
It would treat attention as infrastructure.
It would treat remote work as a serious operating mode for knowledge work.
It would treat trust as something to be designed, not assumed.
It would treat memory as an operational asset.
It would treat human oversight as architecture, not PR.
It would treat mental health and daily life not as softness, but as part of the decision environment.
And it would let organizations become more autonomous without becoming less human.

That is the world I imagine.

Not a world where humans disappear.

A world where systems finally become worthy of the humans inside them.

Sibling essay (same coordination thesis in a different room): Dating, the Price of Anarchy, Proxenio, and the Missing Operating Layer.

Sources, signals, and further reading#