the evidence behind the rise of the agent-native company: definitions survey, company classification, spend-vs-redesign data, and the historical record. confidence flags: [v] verified against a fetched source · [l] likely, strong secondary reporting · [u] unverified, usually the company's own claim.
metr's 50% time horizon measures the length of task — calibrated to how long a human professional takes — at which a model succeeds half the time. it is the closest available measure of how long an agent can work unsupervised, and it is the load-bearing evidence that an agent-native company only recently became possible.
| date | model | 50% time horizon | flag |
|---|---|---|---|
| 2023-03-14 | gpt-4 | 3.99 min | [v] |
| 2024-06-20 | claude 3.5 sonnet | 11.4 min | [v] |
| 2024-12-05 | o1 | 38.8 min | [v] |
| 2025-02-24 | claude 3.7 sonnet | 60.4 min — first past an hour | [v] |
| 2025-08-07 | gpt-5 | 203 min (3.4 hr) | [v] |
| 2025-11-24 | claude opus 4.5 | 293 min (4.9 hr) | [v] |
| 2025-12-11 | gpt-5.2 | 352 min (5.9 hr) | [v] |
| 2026-02-05 | claude opus 4.6 | 719 min (12.0 hr) | [v] |
what it measures, precisely: success rate on software- and research-engineering-shaped agentic tasks in a standard scaffold. it is not wall-clock runtime and not a reliability guarantee — 50% is a coin flip. metr states measurements above 16 hours exceed the current suite, and the fit is revised between suite versions (opus 4.5 appears as 320 min in the 1.1 blog and 293 in the raw file), so treat any single figure as ±roughly a factor of two.
the doubling is accelerating. long-run (2019–2025) the horizon doubles about every 7 months; measured from 2023 it is ~131 days; from 2024, ~89 days. metr's march 2025 paper reported models succeeding on <10% of tasks taking humans more than ~4 hours — fourteen months later the 50% horizon sat at twelve.
| per-step reliability | 10 steps | 50 steps | 100 steps |
|---|---|---|---|
| 90% | 34.9% | 0.5% | 0.003% |
| 95% | 59.9% | 7.7% | 0.59% |
| 99% | 90.4% | 60.5% | 36.6% |
| 99.9% | 99.0% | 95.1% | 90.5% |
plain arithmetic (p^n), but it explains the discontinuity: a per-step improvement that looks incremental on a benchmark shows up in practice as a step change in how long a leash an agent can be given. the rigorous version is toby ord's is there a half-life for the success rates of ai agents? (arxiv 2505.05115, may 2025), which fits metr's data with a constant hazard rate per unit of task time — i.e. every agent has a characteristic half-life. note: we found no published primary claim that models "crossed a reliability threshold" at a specific per-step figure; the 99%/99.9% framing is illustrative, not measured. the measured claims are the horizon curve and ord's fit.
the honest limit on all of this: the horizon suite is coding-shaped. on general computer use the picture is much weaker — epoch's analysis of osworld found ~45% of its tasks need little or no gui interaction and ~10% contain outright errors, concluding it "does not definitively demonstrate reliable general computer use" [v]. osworld 2.0's long-horizon workflows (skilled humans >1 hour) see frontier agents finish about 20% end-to-end. the capability claim we make is about the coding- and ops-shaped work agent-native companies start with, not all knowledge work.
"we could not find published cases" is a true and checkable statement; "none exist" is not. we make the former claim.
the term has no formal definition. surveying current usage:
| source | their definition | framing / what it misses | flag |
|---|---|---|---|
| crv | "ai isn't a feature or an enhancement, but the architectural foundation on which the entire product depends." test: "if you remove the ai, does the product cease to function? not degrade… stop working entirely." | product architecture. distinguishes ai-native / ai-first / ai-enabled. uses "agents within guardrails" but never "agent-native" | [v] |
| ibm | "designed from the ground up with ai as a core component, not bolted on later as a mere feature"; ai "shapes architecture, decision-making, user experience and the entire system lifecycle from the outset" | system lifecycle. ai-native vs ai-augmented | [l] |
| hbs online | "built from the ground up with ai at its core," with "ai embedded in every stage of the organization, from r&d to marketing, customer engagement, and human resources" | whole organization, no operational test | [l] |
| kore.ai | "ai is the architectural foundation for how the organization understands context, acts, learns, and improves"; "built around intelligence rather than static workflows" | names "retrofitting ai into human-designed workflows" as the #1 ai-native mistake — closest to our thesis | [l] |
| mckinsey | doesn't use "ai-native." their term is the agentic organization: humans and agents "side by side at scale at near-zero marginal cost"; "not automation as usual on top of existing processes — it's a redesign of end-to-end processes with humans 'above the loop'" | the closest analyst cousin to agent-native; org-redesign framing | [l] |
| sequoia / redpoint (2026) | "from hierarchy to intelligence": the org chart's information flow replaced by a continuously updated "company world model" built from machine-readable work artifacts | org structure, not product | [l] |
three distinct framings — product architecture (crv, kore.ai), system lifecycle (ibm), whole organization (hbs, mckinsey, sequoia). none requires workflows designed for agents, and none treats prior ai-native investment as an obstacle to reaching that state. that is the gap this paper fills.
no consensus definition and no major-analyst adoption — gartner and mckinsey both say "agentic," not "agent-native." what is genuinely new here is defining it as the internal operating-system end state, and arguing that ai-native investment obstructs reaching it.
| rung | definition | test | where the walls are |
|---|---|---|---|
| 1 · legacy | the pre-ai company. humans execute workflows; software records them. | ai plays no structural role | departments own workflows; knowledge lives in people |
| 2 · ai-enhanced | ai and agents embedded into the legacy structure: copilots, ai product features, agents automating steps inside existing workflows. | remove the ai and everything still works | same walls as legacy, now cheaper to operate |
| 3 · ai-native | designed around ai from day one. ai is connected across the company's systems and drives whole workflows across them — but the systems are still silos built for humans. | remove the ai and there is no company | silos remain; the ai traverses human-built interfaces, approvals, and duplicate records |
| 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 by design |
placement from public reporting as of aug 2026. flag: [v] independent reporting · [m] company marketing or self-reported.
| company | rung | evidence | flag |
|---|---|---|---|
| most of the fortune 500 | 1→2 | gartner: <5% of enterprise apps had task-specific agents in 2025, forecast 40% by end-2026 — agents embedded in existing applications | [v] |
| microsoft copilot buyers | 2 | 20m+ paid seats, ~70–90% of the fortune 500 licensed — but independent surveys put weekly active use at 20–30% of purchased seats. adoption numbers measure procurement, not transformation | seats [m], usage gap [v] |
| salesforce (agentforce) | 2 | ~$800m arr, 29,000 deals; benioff attributed ~4,000 support-role reduction to agents handling ~half of interactions. community analysts dispute depth of adoption | [m]/[v] |
| klarna | 2 — the cautionary tale | 2.3m conversations in month 1; "work of 700 fte agents" (a workload-equivalence calc — hires avoided, not layoffs), ~$40m→$60m savings claimed. may 2025 reversal: rehiring humans; ceo: cost was "a too predominant evaluation factor… what you end up having is lower quality" | claims [m], reversal [v] |
| jpmorgan | 2→3 | llm suite to 200,000 employees in ~8 months; 100+ genai solutions in production; ~$2b/yr ai spend inside a $19.8b tech budget; stated goal "every process powered by ai agents" | [v] |
| walmart | 2 | consolidating dozens of ai tools into four "super agents" layered onto existing retail systems | [m] |
| moderna | 2→3, structurally novel | merged hr and it functions to manage a workforce it calls "part human, part gpt"; 750 employee-built gpts in 2 months → 3,000+ a year later | [v]/[m] |
| goldman sachs | 2, near-4 signal | gs ai assistant to 10,000 employees; piloting an ai coding agent described as "employee #1" of a hybrid workforce across a 12,000-person eng org | [v] |
| block | 3→4 — the clearest large mover | built + open-sourced the goose agent framework (apache 2.0, co-developed mcp with anthropic); self-reported 40% more code shipped per engineer in ~6 months and 8–10 hrs/wk saved per employee; feb 2026 cut 4,000+ jobs (~40% of headcount); jul 2026 launched buzz, a collaboration platform for humans and agents together, itself built using agents. numbers are self-reported/earnings-call, not third-party audited | goose [v], buzz [v], productivity + cuts [l/m] |
| shopify / duolingo | 2→3 (policy, not architecture) | ceo memos making ai use a baseline expectation and requiring teams to prove ai can't do a task before requesting headcount; ai use in performance reviews | [v] (memos self-published) |
| notion, canva, duolingo (product) | 2 | ai bundled into pre-llm products. remove the ai and the product still works — textbook rung 2 | [v] |
| cursor / anysphere | 3 | $500m arr jun 2025 at ~60–150 employees; $1b arr nov 2025; 1m+ dau. product cannot exist without ai; org structure conventional as reported | [v] |
| midjourney | 3 | $200m+ arr at ~40 employees (2023), zero external funding | [v] (sacra est.) |
| lovable | 3 | $100m arr in 8 months at ~45 employees — fastest $0→$100m recorded | [v] |
| elevenlabs | 3 | $330m arr 2025 at ~400 employees; $500m arr apr 2026 | [v] (sacra) |
| harvey | 3 | $100m arr aug 2025 at ~350 employees; $190m arr jan 2026; 1,500+ customers | [v] |
| perplexity | 3 | ~$200m arr sept 2025, >$450m mar 2026 — but headcount 500→1,400 in a year, diluting the lean-team story | [m]/[v] |
| sierra | 3, sells rung-4 | $100m arr in 7 quarters; autonomous agents sold priced per resolved conversation, not per seat — buying an outcome, not a seat | [v] |
| 11x | claimed 4 — not supported by independent reporting | marketed "ai employees" running sales; techcrunch reported it displayed logos of companies that were not customers, and quoted zoominfo saying the product "performed significantly worse than our sdr employees"; the ceo stepped down may 2025 | [v] — reporting verified; claims unverified |
| artisan | claimed 4 — marketing outruns the disclosed evidence | "stop hiring humans" billboards; industry reporting has placed its gross retention below 60%, a figure the company has not publicly confirmed or disputed to our knowledge | claims [m], reported figure [l] |
| enterprise-scale agent-native | 4 | no verified example found | — |
the finding. no well-known company passes a verified rung-4 test, and the loudest claimants are precisely the ones independent reporting punctured. the credible near-rung-4 evidence is quieter and organizational rather than technical: moderna's hr+it merge, goldman's "employee #1," sierra's per-resolution pricing. an org chart, a job title, and a pricing model.
| company | employees | revenue / users | ratio | year |
|---|---|---|---|---|
| instagram (at acquisition) | 13 | ~30m users / $1b price | ~2.3m users per employee | 2012 |
| whatsapp (at acquisition) | 55 | ~465m users / $19b price | ~8.5m users per employee | 2014 |
| midjourney | ~40 | $200m+ arr | ~$5m per employee | 2023 |
| lovable | ~45 | $100m arr | ~$2.2m per employee | 2025 |
| cursor / anysphere | ~60–150 | $500m arr | ~$3–8m per employee | 2025 |
| elevenlabs | ~400 | $330m arr | ~$0.8–1.25m per employee | 2025 |
| harvey | ~350 | $100m arr | ~$0.3m per employee | 2025 |
| perplexity | ~1,400 | ~$200m arr | ~$0.15–0.44m per employee | 2025 |
| meta / google (benchmark) | — | — | ~$1.6m / ~$1.8m per employee | 2023 |
the complication we keep in. the ai-native leaders at their leanest beat big tech's revenue per employee by 2–5x — but every one that scaled into enterprise sales reverted toward ordinary saas economics. on current evidence the tiny-team advantage is a property of product architecture, not of the operating model. closing that gap is what agent-native design has to prove.
the synthesis: record spend × an embed-in-existing deployment pattern × a 21% redesign rate = a market paying to make its current shape permanent. the 50%-of-agents-in-silos figure is the rung-3 ceiling stated numerically.
| wave | the bolt-on | the native move | the evidence |
|---|---|---|---|
| internet (1995–2005) | add a website; borders outsourced its own site to amazon (2001–08), booking online as "loss reduction" | a company impossible without the internet | 2004: blockbuster ($6.1b revenue, 9,000 stores) launches a bolted-on site; netflix does $500m. blockbuster ch.11 in 2010 with ~$1b debt; borders liquidates 2011. us e-commerce ~1% of retail (2000) → 16.9% (q1 2026) |
| mobile (2007–2018) | add an app to the desktop product | impossible-before-the-phone companies | nokia ~50% smartphone share (2007) → ~3–5% (2013); blackberry ~20% (2009) → <1% (2016). nyc taxi medallions: peak ~$1.05m individual / $1.32m corporate (nov 2013) → ~$160–170k by 2018, ~$93k by q2 2022 (~90% off peak) as uber built the business around the rider's phone [l — bloomberg data via secondary]. kodak ch.11 jan 2012 (peak ~140,000 employees); three months later instagram sells for $1b with 13 employees; whatsapp $19b with 55 |
| cloud (2008–2018) | lift-and-shift the datacenter | re-architect around elasticity | siebel: ~45% of crm (2002), ~$2b peak revenue, sales −22% in 2002, acquired by oracle sept 2005 for ~$5.85b; cloud-native salesforce (1999, "the end of software") #1 in crm by idc share since 2013 [l]. netflix: aug 2008 → jan 2016, seven years, a full rewrite. "moving as-is means you end up moving all the problems and limitations of the data center along with it." gartner: 45% of lift-and-shifters overspend by 70% within 18 months; rehosted workloads run ~40% more expensive than refactored |
| ai / agents (2023–) | embed agents in existing workflows | design the company for agents | in progress — this paper's thesis |
the structural constant: each wave's bolt-on produced real first-wave efficiency and acted as a commitment device against the native redesign. roughly half the year-2000 fortune 500 no longer exists independently; average s&p 500 tenure fell from 33 years (1965) to a projected ~14.
davenport, harvard business review, july–august 1998 (3,891 citations): unlike bespoke systems built around a company's requirements, enterprise systems "impose their own logic on a company's strategy, culture, and organization." you redesign the business around the system, or the project fails. hammer & champy had already defined the program in 1993: "fundamental rethinking and radical redesign of business processes." fewer than 45% of bpr attempts hit their goals — the redesign was the hard part, which is why installing the software without it was hopeless.
| case | year | damage | cause |
|---|---|---|---|
| foxmeyer drug | 1994–96 | $65m project; the $5b-revenue company bankrupt in 1996 | order throughput collapsed from 420,000 to 10,000 orders/night; $34m inventory losses |
| hershey | 1999 | $112m rollout; ~$100m of halloween orders undeliverable; q3 sales −12.4% | big-bang sap+manugistics+siebel compressed from 4 years to 30 months to beat y2k — testing and training were the casualties |
| nike / i2 | 2000–01 | $100m sales shortfall; stock −20% in a day | demand-planning mis-forecasts; >10–15% customization against vendor guidance |
| lidl (elwis) | 2011–18 | ~€500m sunk against a €201m budget; abandoned after 7 years | refused to adopt sap's retail-price inventory valuation, customized the package to mimic its legacy process until it collapsed — reverted to the in-house system. the purest install-without-redesign failure |
| revlon | 2018–19 | ~$64m of shipments unfulfilled; three shareholder class actions | s/4hana go-live crippled a plant; first erp failure to trigger securities suits over the implementation itself |
lock-in economics: 55–75% of erp projects fail to meet objectives; 55% run over budget, 68% over schedule. maintenance runs ~20%/yr of license and escalates ($1m/yr becomes $1.63m by year 10). systems are kept 10–15+ years. the structural definition worth borrowing: lock-in happens when the cost of leaving exceeds the cost of tolerating the system's limitations.
the organizational-design questions this raises — what a department is, what a manager is when coordination is a protocol, what accountability means when the workforce is agents — are the subject of the next report.
method. three parallel research passes (definitions + primary stats · company evidence · historical waves and erp lock-in), each required to attach a source url and state what every number actually measures. confidence is flagged per row rather than averaged. several publisher sites (ibm, hbs, gartner, mulesoft, mckinsey) block automated fetching, so their language comes from search snippets and syndicated mirrors and is marked [l]; only pages fetched directly carry [v].
known weak points, stated rather than hidden: the gartner agent-software forecast ($206.5b 2026 → $376.3b 2027) is the least well-sourced figure in the set and needs one primary check before print; the zapier and deloitte survey numbers came via secondary posts; "roughly half the year-2000 fortune 500 is gone" is ubiquitous but loosely sourced — we present it as "roughly half," never as a census.