the technology now exists to build companies where ai agents are not tools employees use, but the operating system the company runs on. but where are the companies building this way?
read the full research report →
until very recently, "ai does the work" was a category error. a model that could answer a question could not hold a job, because holding a job means working unsupervised for hours, across tools, recovering from your own mistakes — and models fell apart within minutes.
that constraint broke, and it broke recently. metr measures the length of task an ai can finish on its own, calibrated against how long a human professional takes. in march 2023, gpt-4 could handle about four minutes of work. in february 2025 a frontier model crossed one hour for the first time. by november 2025 it was about five hours, and in february 2026, twelve.
the constraint was never intelligence. it was duration. an agent that loses the thread after four minutes needs a human between every step, so it can only ever be a tool. an agent that stays coherent for twelve hours can own the step — and owning steps is what an employee does. the trend line is also getting steeper, not flatter: the long-run doubling time is about seven months, but measured from 2024 it is closer to three. and twelve hours is a reading, not a ceiling — every point on that line looked like the horizon from where it stood.
only very recently — likely this year — has building a company this way become genuinely possible.
we went looking — the analyst definitions, the vendor lists, the loudest public claims — and could not find a company at any real scale that has restructured its operating model around agents. of the thousands of vendors describing themselves as agentic, one analyst count reckoned only about 130 were real, the rest "agent washing" — rebranded chatbots and rpa.
the closest is probably anthropic, which reports more than 80% of the code merged into its production systems is now written by claude, with engineers moved into deciding what to build and reviewing what comes back — an agent-native engineering organization, with the same people-own-agents shape spreading into its other functions. an agent-native function, on its way to an agent-native company.
the sharpest evidence of the gap is who is hiring. the companies selling the agent workforce — sierra, decagon, legora — are staffed conventionally and recruiting humans at speed, a hundred-plus open roles apiece, and none of them post roles whose job is supervising a fleet of agents. the agent-native company is currently a product being sold, not an operating model being run.
the tell to watch is the agent manager — a person hired to direct and review agents rather than do the work. the role now formally exists: harvard business review defined it in february 2026 with the coo of salesforce's agentforce platform as co-author, and salesforce and servicenow have begun posting the title. a company hiring agent managers is restructuring human roles around ownership of agents — the first structural signal, and it is brand new.
that gap is the subject of this paper: what the thing actually is, why almost nobody has built one, and why we think the companies rushing to spend their way past it are making it harder, not easier, to get there.
in an ai world, every company sits somewhere on a ladder of four rungs — defined not by how much ai it buys, but by what its structure assumes. each rung has a simple test:
| rung | what it is | the test | examples |
|---|---|---|---|
| 1 · legacy | the pre-ai company. humans execute workflows, software records them. | ai plays no structural role | most of the economy today |
| 2 · ai-enhanced | ai and agents embedded into the legacy structure: copilots, ai product features, agents automating steps inside existing workflows. same company, faster. | remove the ai and everything still works | salesforce + einstein, notion ai, duolingo max, jpmorgan's llm suite, klarna |
| 3 · ai-native | designed around ai from day one. ai drives workflows across the company's systems — but the systems are still silos built for humans. | remove the ai and there is no company | cursor, perplexity, harvey (product layer — ops claims mostly unverified) |
| 4 · agent-native | workflows designed for agents end to end on one operating system. agents are the workforce; humans direct, review, and are accountable. | work flows through one agent os; output scales without headcount | none verified at enterprise scale; a handful of small companies, ours included |
klarna shows the rung-2 ceiling: its ai assistant did the work of ~700 support agents, and by 2025 the company was rehiring humans — its ceo conceding the result was "lower quality." it had automated the old workflow rather than redesigned the work.
the line between rung 3 and 4 is the one most people don't see. the assumption everywhere is that ai-native is the end state — that a company designed around ai is the finished form. it isn't. the best ai-native operation today has ai driving workflows across its finance system, crm, and data warehouse in a single run — but it spends its day logging into systems built for human hands, waiting on approval chains built for human managers, translating between silos. a brilliant operator of infrastructure never designed for it. the next section is about what sits past that ceiling.
some companies claim rung 4 today; the loudest don't pass the test. and being built by ai doesn't make a company agent-native — the test is whether the business runs on agents, not whether ai wrote the code.
most people are not thinking about the last rung at all. the corporate conversation is about moving from 1 to 2, and the ambitious edge of it from 2 to 3. the clearest large company even pointed at 4 is block: jack dorsey built the open-source agent framework goose, drove it across the whole company — a 40% increase in code shipped per engineer within six months — then cut over 40% of headcount and shipped buzz, a collaboration platform built for humans and agents to work in together. dorsey's own framing: "i don't think we're early to this realization. i think most companies are late."
very little has been written on this, so this section is firsthand — from inside a company that operates this way.
agents hold roles, not tasks: a defect agent finds, fixes, and ships. a pm agent drives plans. a cfo agent tracks spend against caps. all of them read and write one operating system, so work moves from found to fixed to shipped to accounted for without crossing a wall. the substitutions underneath: agents instead of departments, goals instead of workflows, context instead of documentation, autonomous execution instead of handoffs, human judgment instead of human labor.
agent-native doesn't mean building everything yourself — off-the-shelf tools where they're good enough, replaced later or never; that's build-versus-buy, not a purity test. it cuts the other way too: a small agent-native company can grow a function like finance in context and may never need a finance system at all. the system of record shrinks to what genuinely needs recording.
and the "agents replace employees" framing misses the real change: the role of the human. in a legacy company, an employee owns a slice of a workflow. in an agent-native company, the workflow belongs to the operating system, and the human's job becomes direction, review, and accountability for the agents doing the work. you stop being a step in the process and become responsible for its outcomes.
which raises a question we'll leave open: should agents be on the org chart? our working theory is no — an org chart is still made of people, because accountability has to land on a person. what changes is what a role is: a person is assigned ownership over agents, an agent is responsible for the work, and the human is accountable for it. the org chart stays human; the roles inside it get restructured around agent ownership. we plan to write that up properly, from our own operating experience, as a future post.
internet — most companies added websites. in 2004 blockbuster — $6.1b of revenue, 9,000 stores — launched a website bolted onto its store network; netflix, built around the internet, did $500m that year and blockbuster was bankrupt by 2010. borders went further and outsourced its own website to amazon. it liquidated in 2011.
mobile — most companies treated the phone as a smaller screen for the website. new york's taxi industry treated it as an accessory to the dispatch system it already had; uber built the whole business around the device in the rider's pocket. a nyc taxi medallion peaked at about $1.3m in 2013 and was worth roughly a tenth of that within a decade. instagram — built only for the phone, 13 employees — sold for $1b in 2012.
cloud — siebel owned crm: roughly 45% of the market and ~$2b of revenue at its peak. salesforce launched in 1999 as pure cloud under the slogan "the end of software"; siebel's sales fell and oracle absorbed it in 2005. the rare incumbent that made the native move, netflix, spent seven years rewriting for the cloud rather than lifting-and-shifting — because moving systems as-is means "you end up moving all the problems and limitations of the data center along with it."
| wave | the bolt-on move | the native move | who won the decade |
|---|---|---|---|
| internet | add a website | be impossible without the internet | the natives |
| mobile | add an app | be impossible before the phone | the natives |
| cloud | lift-and-shift | re-architect around elasticity | the natives |
| agents | embed agents in workflows | design the company for agents | in progress |
the pattern is precise: the bolt-on crowd captures the first wave's efficiency; the natives capture the next decade's market. nobody calls themselves "internet-enabled" anymore. everyone assumes this wave's end state is ai-native — but ai-native is this wave's website.
the argument for agent-native isn't ideological. it's what happens to output per person at each rung — and what happens to it again every time a new model ships.
take one use case: conceiving and shipping a new software product — not a toy, a real one, with a backend, an automated data pipeline, and its own decision logic.
| rung | the same build | who + how long |
|---|---|---|
| 1 · legacy | every system hand-built or bought and integrated; work moves through specs, tickets, and handoffs | a product team — pm, design, several engineers — and a quarter or more |
| 2 · ai-enhanced | the same team, the same handoffs, faster keystrokes | the same team, weeks shaved off |
| 3 · ai-native | ai drives the build, but through tools and approval chains made for humans | a small team, weeks |
| 4 · agent-native | agents hold the roles; the human directs and reviews | one person, about a week |
the last row is not hypothetical. we recently took a second product from idea to working system in about a week, one person directing: the full backend, the automated data collection, the optimization engine that ranks what it collects — a build meaningfully more sophisticated than what a funded team would have shipped in a quarter a few years ago. the binding constraint on that week was not people, skill, or hours — it was tokens: how much model capacity the operating system could spend. that is a constraint that gets cheaper every year, on a curve nothing human-shaped follows. and the operation this site reports on tripled its weekly throughput over seventeen weeks while its defect rate halved — quality and speed normally trade against each other; here both came from the operating system getting better at catching itself.
the curves diverge because each design converts a different share of every new model release. a legacy company converts none of it. an ai-enhanced company converts a sliver — the human is still in the loop of every unit of work, so the gain caps at how much faster a human-paced workflow can run. an ai-native company converts more, until it hits the silos. an agent-native company converts nearly all of it, everywhere at once: the same operating system runs every function, so when the model underneath gets a longer horizon, the whole company gets it in an afternoon — and the operating system itself keeps improving between releases.
that is why rung 4 is not one more increment. rungs 2 and 3 make the existing company faster, and the gains show it — tens of percent. rung 4 changes what output scales with: in a legacy company output scales with headcount; in an agent-native company it scales with how good your operating system is. the gap to even the best rung-3 company is measured in multiples, not percent — and it widens with every release.
while the capability sits available, enterprises are in the largest technology spend cycle since cloud — and almost all of it is going into embedding ai, increasingly agents, into workflows that already exist. an agent in the ticketing queue, in the erp approval chain, drafting inside the crm.
that is the software-upgrade mistake: treating ai as a new layer on the company rather than a new foundation for it. short term it produces real value — costs drop, cycle times drop, the pilot dashboards look great. long term it is a trap, and the trap has three jaws.
first, it doesn't move the number that matters. mckinsey found that of ~25 attributes tested, redesigning workflows had the single biggest effect on whether a company saw ebit impact from ai — and only 21% of gen-ai-using organizations have fundamentally redesigned any workflow. most of the spend is buying speed on processes designed for a pre-ai world.
second, it hardens what should be deleted. every dollar spent wiring agents into the current workflow makes that workflow more load-bearing — and deleting it is exactly what the next stage requires.
the enterprise that just spent $40 million putting agents inside the hris didn't move toward the second picture; it bought insurance against ever getting there. the investment, the vendor contract, and the retrained staff are now all reasons not to delete the workflow. lock-in isn't a side effect of the spend — it is what the spend purchases.
third, the dependency compounds faster than erp's did. thomas davenport wrote the sentence that matters in 1998, about erp: enterprise systems "impose their own logic on a company's strategy, culture, and organization." swap in "agent platform" and it publishes today unchanged. as one analyst puts it, agentic lock-in "is more durable than api lock-in because it accumulates at multiple layers simultaneously": the model becomes interchangeable, but the data, the orchestration logic, and the agent runtime do not.
erp already ran this experiment in both directions. lidl spent seven years and roughly €500 million forcing sap to mimic the legacy inventory process it refused to change, then abandoned the project and went back to its old system — the software worked; the refusal to redesign killed it. erp forced companies to ask "why do we have five systems for this?" the agent era asks the harder one: why do we have a system at all?
the instinct is to transform the company you have. the arithmetic is against it: existing processes carry decades of accumulated exceptions, approvals, and integrations, and rebuilding them for agents means renegotiating every one. a competitor who never had them pays none of that cost.
the more promising shape: build the agent-native company alongside the legacy one, and migrate work into it. let the existing business fund the work. somewhere beside it, start from the founding question — if this company were started today with agents as the workforce, what would it build? — and let a small unit operate that answer on real work, not as a lab. move functions across as the new side outperforms the old.
the argument for starting now is that nobody has finished. begin today and the field is roughly level. wait, and the eventual competitor is a company whose cost of serving a customer falls every year while payroll-driven costs rise — the great repricing modelled that crossover, and once the curves cross they don't cross back.
erp digitized the enterprise. agents will reorganize it.