a measurable definition of the agent-native company, a benchmark for verifying the claim, the maturity ladder inside the category, evidence-based placements of the public claimants, the capability and enabler timeline — and the benchmark's first live run. companion to the blog post; extends no. 003.
confidence flags throughout: [v] = primary source fetched during this study · [l] = secondary or corroborated via search · [estimated] = our own judgment-involved measurement, method stated.
five research lanes, run august 2026: existing frameworks · company evidence · capability timeline · self-measurement · instrument precedent. every load-bearing figure carries what it measures, its source, and a confidence flag; open verification debt is listed in section 09.
observer bias, stated plainly: the authors operate the company graded in section 07 and originated the benchmark. controls: metric definitions pinned before measurement; every number produced by a committed script reading a system of record, re-runnable for any window; classification validated against a hand-labeled sample with error bars published; failing and unflattering measures published; the same confidence flags applied to ourselves as to every company in section 04.
what this study cannot tell you: no control company exists — a conventional twin given the same objectives is the right experiment and is not feasible for a one-person operation; the substitutes are longitudinal self-comparison and a published benchmark others can run. and the operation graded here is small, so nothing in section 07 demonstrates the benchmark at enterprise scale.
ten frameworks surveyed. the finding, verified framework by framework: everything in circulation counts agents, describes capability tiers, or checks governance boxes. nothing measures what share of completed output was agent-executed end to end, verified from systems of record.
| framework | unit graded | measures or describes | the specific gap |
|---|---|---|---|
| microsoft agentic maturity model (100–500) [v] | the adoption program | describes — cmm-style practice statements | its top level, "agent-first enterprise," is awarded with zero work-share evidence; a company could clear it with strategy docs and governance boards while humans perform every execution step |
| salesforce agentic maturity model (l0–l4) [l] | the deployed agent's capability | describes + checklists | one l4 multi-agent demo makes a "level 4 company"; it is also an agentforce adoption roadmap |
| gartner hype cycle + maturity toolkit [l] | the market / self-reported practice | analyst judgment + survey | "deployed an agent": 17% yes — a yes/no that can't distinguish a pilot chatbot from an agent-run company |
| iiq, arxiv 2605.14455 [v] | ai usage intensity | genuinely measures (0–1000, continuous) | weights input proxies — novelty-weighted token stock — demonstrated only on synthetic scenarios; a company burning tokens while humans perform every consequential step scores high |
| alchemy crew har / ami [v] | agents-per-human, per function | hybrid; no published method | counts entities, not output — 200 trivial agents per human outscores 3 agents executing 90% of a workflow |
| agentuity, "the agent-native company" pts 1–2 [v] | the company, qualitatively | describes, vividly | the best definition in circulation ("take away the ai and the company stops functioning") with zero operationalization — no metric, threshold, or evidence standard |
| knight institute autonomy levels l1–l5 [v] | one agent–user interaction contract | taxonomy + proposed certificates | says nothing about an organization; a company could run fifty l5 agents on 1% of its work |
| anthropic economic index [v] | conversations with one vendor | measures, at scale | the best public trend line on delegation (directive share 27%→39% across 2025 [l]); no org identity, no completed work items |
| openai gdpval [l] | single work deliverables | measures — blind expert grading | task ≠ workflow ≠ standing responsibility; the strongest evidence individual deliverables reach parity (~48% win+tie at ~100x speed and cost [l]) |
| mckinsey trust maturity / state of ai [l] | governance posture, surveyed | self-reported scores (avg 2.3/5) | its own thesis — "redesign the work" — concedes the gap its instrument can't measure |
the term itself is overloaded at least four ways in current use: a company run by agents (agentuity's sense, and this report's) · a software architecture where agents are first-class users of the product (builder.io, every — "agent ui parity") [l] · cloud infrastructure re-architected to run agents (alibaba's "agent native cloud," waic 2026) [l] · products built for agent customers (per-outcome pricing, mcp-first surfaces) [l]. no current sense carries a measurable, auditable definition of an organization run by agents. that is the gap this report fills.
| # | measure | definition | source | computation | gaming risk / limits |
|---|---|---|---|---|---|
| 1 | agent work share | completed work items agent-executed / all completed, in window | work tracker + version control | classify every item against 02.1's definition; automated attribution is a proxy — validate on a random hand-labeled sample (≥50 items), publish the agreement rate and carry it as the error bar | item-splitting inflates the numerator — require uniform item definitions; exclusions are named rules, applied uniformly, counted |
| 2 | human intervention rate | human execution steps / all execution steps; direction, review, approval logged separately | labeled sample | from the sample: items where a human performed an intermediate step. report the CEILING as the measure — an item whose records do not say who performed a console or device step counts as human-touched until instrumentation says otherwise; report the confirmed floor alongside it, never instead of it. note the population: work items, not code changes; the two are different denominators and must not be blended | step granularity is manipulable — pin the step definition before measuring; human execution is observed through artifacts, and consoles and devices produce few, so report the reading as a floor with its upper bound |
| 3 | initiative share | completed work originating from agent-identified opportunities / all completed | tracker creator + actor fields, sample-validated | machine-filed work is the automated floor; agent-noticed-but-window-filed work requires the labeled sample | agents auto-filing trivial records inflates it — count only work that shipped |
| 4 | wall count | human-mediated context transfers between functions, per end-to-end workflow | workflow tracing, case-documented | walk each core workflow; a human typing "go" is a gate, not a wall | qualitative — publish the traces; verify every "a mechanism carries this" claim against an artifact that exists |
| 5 | headcount elasticity | Δ output vs Δ headcount, trailing window | tracker + payroll | weekly completed-item series against headcount series | output-metric choice; publish the series with every known discontinuity dated (model upgrades, convention changes, product launches) |
| 6 | marginal cost mix | compute vs payroll in the next unit of output | billing + time records | tokens + infrastructure for a representative shipped unit vs the human hours on it | unit-selection bias — state the selection rule |
| 7 | reorg cost | elapsed time and change size to stand up or reshape a function | dated configuration diffs | decision timestamp → in-force timestamp | anecdotal by nature — date-stamped artifacts required; case choice is judgment and says so |
| 8 | function autonomy + census | every business function graded: manual · assisted · task · workflow · standing · self-directed | role configs, triggers, tracker | one line of evidence per grade; standing requires an enumerable trigger that fired unprompted in the window | self-scoring drift — evidence per cell; functions consolidated at the business level (engineering absorbs review, testing, observability), sub-function detail kept beneath |
| 9 | organizational time horizon | human-equivalent task duration completed reliably, across operations | labeled sample, duration-bucketed | success rate per duration bucket at 50% and 80% thresholds | durations are estimates unless measured on humans; the population is self-selected — a company only attempts what it believes agents can do — so readings are a ceiling on attempted work |
the autonomy grades map to the knight institute's user-role levels at the function grain: at what level does the human sit for this function, for what share of its work.
three probes anchor the benchmark because they are hard to fake; an examiner runs them, the company doesn't self-report them. shutdown: pause every standing trigger for 24 hours (the products stay up — pausing them measures uptime, not autonomy); measure what stops arriving. autonomous sprint: a standing objective, five days, human answers only safety-gate questions; measure what got discovered, built, shipped, escalated — and how much of it the human would have chosen. saas-deletion (thought experiment): could the agents recreate the workflows the saas provided? these are proposed audit procedures, not stunts; a small operator monitoring its agents at the level current reliability deserves may reasonably decline to run the first two, and section 07 does.
for companies that cannot be audited: hiring pattern (what roles are posted) · org-chart titles (agent supervision) · pricing model (per outcome vs per seat) · the build test (idea → shipped product: people × time) · published work-share numbers, and whether the methodology is disclosed.
the benchmark presumes four data classes: one work graph (every task, defect, decision in one tracked queue, agent- and human-written) · one context layer (conventions, constraints, lessons, role definitions stored where agents load them) · one execution record (every change attributed, approval trail attached) · standing triggers (enumerable schedules and event hooks). an agent-native company holds these by construction, so measures 1–9 are queries; a conventional company cannot produce them without months of instrumentation. the asymmetry is itself a test: "we cannot produce these numbers" is the rung-3 diagnosis in one sentence.
no grading floor exists anywhere; this is a proposal, open to argument, with each line's reasoning:
4.2 adds (hypotheses): meaningful initiative share; review moved from outputs to exceptions; multi-day 80%-reliability horizons. 4.3 adds: enterprise scale, customer-facing, external audit, materially better economics per human hour.
two axes underneath rung 4 — coverage (how much of the company runs on the agent os) and autonomy (what humans still touch: every output → exceptions → policy). the ladder 4.0 → 4.3 is the diagonal. boundary cases no. 003 left open, resolved by the benchmark:
placed on the public-evidence tier. negative characterizations are attributed to named reporting throughout.
| company | rung (evidence) | evidence | flag |
|---|---|---|---|
| anthropic | 4.0 — engineering | "more than 80% of the code we merge into anthropic's codebase was authored by claude," with a stated method — share of lines merged to production attributable to claude (may 2026); typical engineer merging 8x the code of 2024 | [v] |
| block | 3 → 4.0 | q1-2026 shareholder letter: builderbot makes 15% of production code changes "nearly fully autonomously, with people only intervening to make the final decision to push to production"; ai reviewed >90% of change requests; 200k operations/day; code changes per engineer up 2.5x since january; incident rate after changes down 70% year on year. the feb-2026 ~40% workforce reduction is attributed to ai by the company; named reporting calls the ai contribution "hard to isolate" | [v] letter · [l] cuts |
| salesforce | 2–3 internally; claims more | its ceo told bloomberg's the circuit (june 2025) "ai is doing 30% to 50% of the work at salesforce" — no methodology; bloomberg's own headline said 30%. meanwhile salesforce formally defined the agentic work unit in its earnings releases — "discrete tasks executed by ai agents in production across the salesforce platform" — 3.8b in q1 fy27, +111% quarter on quarter: a customer-product metric. the company that owns the unit of agent work publishes no number for its own organization | [l] claim · [v] awu |
| basis | 3 | $100m series b at $1.15b (feb 2026); agents used by ~30% of the top-25 accounting firms; an internal team (atlas) builds agents across its own engineering, sales, and talent. its public job postings read like a conventional software company's; no agent-supervision titles found, no internal operating numbers published | [v] |
| maxima | 2–3 | agentic accounting vendor; claims of 95% automation for customers are self-reported and unverified | [l] |
| anthropic–blackstone–h&f–goldman jv | watch | $1.5b services firm (may 2026) incorporated to run client work on claude agents from day one — the purest "build it that way" test at services scale; nothing disclosed yet | [l/v] |
| goldman sachs | 3, entering 4.0 in back office | autonomous claude agents in trade accounting, compliance, onboarding, built with embedded anthropic engineers (cnbc, feb 2026); $150m into the jv above | [l] |
| jpmorgan | 2→3 | plans agents "operating without human intervention for hours at a time" in 2026; 20% private-banking gross-sales lift attributed to ai (cnbc) | [l] |
| microsoft | 2–3; the telemetry landlord | work trend index 2026: active agents in m365 up 15x year on year (18x in large enterprises) — ecosystem telemetry; no self-measurement published | [v] |
| coinbase | 2–3, function-level assist | its ceo (sept 2025): ~40% of daily code ai-generated, target 50% — assist share, explicitly reviewed; the definitional contrast case | [l] |
| sierra / decagon / legora | 2 | sell the agent workforce, staff conventionally at scale — legora's headcount up ~9x since 2023 with 138–189 open roles (revelio data; interpretation ours); sierra prices per resolution (the one public-proxy positive) at a $15.8b valuation | [l] |
| klarna | 2, arc complete | the canonical cautionary case now has an ending: automated the old workflow (work of ~700 agents, 2024), reversed (2025), settled into a hybrid with cost per transaction down 40% over two years — it found the boundary rather than simply failing | [l] |
| shopify / duolingo | 2, policy not architecture | ai-usage mandates; duolingo's partially walked back after user backlash | [l] |
| 11x / artisan | claimed 4; evidence 2 | prior reporting punctured 11x's customer claims (no. 003 sourcing); artisan's founder publicly reversed the "stop hiring humans" slogan (techcrunch, apr 2026) — and its "ceo replaced by ai" story was an april fool's stunt, per the company | [l] · stunt [v] |
the finding: above 4.0, the ladder is empty as measured — occupied only as claimed. three organizations now publish an agent work metric (anthropic, block, salesforce), each real and each differently scoped — two to engineering, one to product usage. nobody publishes an org-level, cross-function agent work share, an intervention rate, or an initiative share. the disclosures that would move any company up-rung are exactly the benchmark's measures.
the two-frontier picture (metr, raw data [v]): at 50% reliability the frontier completes tasks of ~12 human-equivalent hours (claude opus 4.6, feb 2026; confidence interval ~5–60h; metr marks measurements above 16h unreliable on its suite). at 80% reliability the same model's horizon is ~70 minutes (ci 27–170). the 80% frontier is doubling roughly every 3.3 months (1.7 min in june 2024 → ~186 min by april 2026, our computation on metr's rows); the 50% frontier every ~4.3 months from 2023 (ci 104–158 days). an organization runs on the 80% number.
the domain gap: on realistic computer-use workflows (osworld 2.0 [v]), the best system completes 20.6% of ~1.6-hour-median workflows end to end, and no model completes anything above ~163 minutes — one to two orders of magnitude behind coding, consistent with metr's measured 40–100x domain gaps. the reliability ceiling: some work needs 98%+ reliability to be worth automating; no current benchmark can fit 99%+ horizons at all [v]. the model of failure: ord's constant-hazard analysis [v] predicts 80% horizons at ~1/3 of 50%; observed ratios are 5–10x wider — today's models are flakier over long tasks than the simple model predicts — though a feb-2026 revision (hamilton, via ord) finds hazard rates decline within a task, mildly favorable for multi-day work.
when (our extrapolation on metr's numbers, their caveats carried): multi-day 80%-reliability horizons — the capability precondition for 4.2 — land mid-2027 at the fast doubling rate, late-2028 at the slow one. 4.3 is gated by rails, not the curve: agent identity and payments already run at volume (a2a: 150+ organizations in production [l]; x402: 169m payments in year one [l]); the audit rail is nascent (aiuc-1 is the first agent-behavior standard [l]; iso 42001 audits the management system around ai, not agent work; no "soc 2 for agents" exists); the legal rail is absent — no settled liability framework and no court precedent found for autonomous-agent harm [l]. earliest plausible verified 4.3: 2028–2030. honest gap: multi-agent orchestration ("agents managing agents") has no metr-grade eval; every number in that area is secondary [l].
three precedents, split by forcing function. cmm (sei, 1987, under dod contract; v1.0 1991): a monopsony buyer demanded a maturity rating, contractors complied, the rating leaked into commercial procurement — and its documented failure mode is maturity theater: organizations building to the checklist, level-5 ratings on failed products [v]. sae j3016 (2014): the l0–l5 vocabulary won when nhtsa abandoned its own competing taxonomy and adopted sae's (september 2016), and insurers and the uk's automated vehicles act followed — but j3016 measures design intent, not safety, and manufacturers self-declare their level, which is how "full self-driving" got marketed on an l2 system [v]. soc 2 (aicpa, 2010, replacing sas 70): the durable trust form — defined criteria + evidence over a period + an independent examiner; the type 1 / type 2 split maps cleanly to "designed agent-native" vs "operated agent-native for a quarter" [v].
the current landscape replays the history: iso 42001 certifies ai governance, not agent work; the big 4 use agents to run audits but attest nothing about agent-native organizations; soc 2 is being visibly stretched over agents (sas 70, replaying); the eu ai act's mostly self-assessed conformity repeats the j3016 flaw at regulatory scale [v/l]. the likely first forcing function is investors — the agent-native claim is made to them before it is made to customers, and no investor we can find publishes agent-native diligence criteria [l].
the composite design lesson, applied in section 02: a j3016-style ladder with soc-2-style evidence requirements, scoring outcomes rather than process artifacts — because cmm's history says a process-theater version will be supplied by an incumbent if a measurement version doesn't exist first.
subject: the authors' company — one person, started march 2026, two consumer products in beta. window: 2026-07-20 → 2026-08-16 (four complete weeks), against a 21-week trend series. method: definitions pinned before measurement; every figure from a committed script reading the work tracker, version-control history across six repositories, and billing records; classification validated on a hand-labeled random sample of 60 completed work items; snapshot inputs checksummed.
| criterion | threshold | reading | grade |
|---|---|---|---|
| agent work share | > 50% | 98.3% of 1,777 completed items [95% ci 91.1–99.7]; conservative reading 91.7% (charges every unattributed console or device step to a human) | exceptional |
| human intervention rate | < 30% | 8.3% (5 of 60 sampled work items) — the reported figure, and deliberately the ceiling: one is a confirmed human execution step (a purchase the founder had to make on a device to reproduce a defect) and four are items whose records do not attribute an outside-the-system step, which we count against ourselves rather than assume in our favour. confirmed floor 1.7%. on a separate population — 25 sampled code changes not attributed to an agent — none were human-written (95% upper bound 13%). direction ran at 3,868 human messages against 184,052 agent tool calls (~48:1). the floor/ceiling gap is an instrumentation gap; logging human out-of-system actions is tracked work, and subsequent readings should report a single number | exceptional |
| standing responsibilities | ≥ 1 per core function | 68 scheduled jobs + 4 event triggers, enumerated from the machines and repositories, spanning every core function | pass |
| function autonomy | core at workflow+ | engineering self-directed · it infrastructure standing · finance ops standing · product workflow · marketing task (census below) | pass |
| human role | zero standing human execution steps | 2 remaining walls (a beta user's feedback arrives by voice to the founder; one class of fix needs a privileged command); 1 confirmed human execution step in 60 sampled items | pass |
| duration | ≥ 4 weeks + 12-week trend | the window is 4 weeks; the operation has run this way since march — 21 continuous weeks, thresholds holding throughout | pass |
| organizational time horizon | reported, not thresholded | completion ~90% from 15 minutes to 3 days of human-equivalent work; self-selected population; no unattended multi-day tasks attempted | borderline |
| function | grade | evidence (one line) | the human's role |
|---|---|---|---|
| engineering — incl. review, testing, observability, the defect lane | self-directed | 68.3% of completed work agent-originated [ci 55.8–78.7]; 3,166 detector records filed and closed in-window with no human in the loop | direction, product calls, the merge word |
| it infrastructure | standing | patrols, watchdogs, health sweeps on schedules | an occasional privileged command |
| finance ops | standing | daily and weekly spend reporting from billing apis, unprompted | approves purchases |
| product — management, design, research | workflow | plans, mockups, research produced end to end in-session; no standing product cadence runs | commissions, decides, approves |
| marketing | task | engagement automation is built but not running as a daily operation in the window — graded on what runs, not what is built | runs it; approves everything outbound |
| legal & compliance | assisted | on-demand counsel workflows; not instrumented | does the filings |
| customer support | assisted | one product files reader reports automatically; the other's beta feedback arrives by voice | is most of the queue |
| sales / bd | manual | does not exist | — |
score: 4.1 — core functions at workflow-to-self-directed with humans reviewing consequential actions; not 4.2 (coordination is agent-assisted, not agent-owned; review has not moved to exceptions), nowhere near 4.3. the functions that grade highest are the most instrumented, which is exactly the bias a self-administered benchmark should be assumed to carry.
two-tier source ledger — fetched-and-verified this study vs corroborated-via-search. primaries fetched [v]: metr time-horizon raw benchmark data and limitations notes; anthropic "when ai builds itself"; block q1-2026 shareholder letter; salesforce q4 fy26 + q1 fy27 earnings releases (agentic work units); zapier ai-vendor survey; microsoft agentic maturity model; microsoft work trend index 2026; agentuity pts 1–2; knight institute autonomy levels (arxiv 2506.12469); iiq (arxiv 2605.14455); alchemy crew har/ami; anthropic economic index (primitives); osworld 2.0 leaderboard; ord half-life (arxiv 2505.05115); epoch on economic-value benchmarks; sei/cmmi histories; sae j3016 histories and nhtsa adoption records; soc 2 histories; eu ai act article 43.
open verification debt before wider distribution (each currently [l], corroborated across multiple secondaries): gartner 17%/42% deployment stats · x402 169m payments · a2a 150+ organizations · aiuc-1 claims · gdpval headline figures · legora/decagon headcount data (revelio) · gartner agent-software spend forecast primary · the bloomberg benioff interview primary (paywalled; wording corroborated across three outlets).
estimation conventions: list-price compute valued at published per-token rates with cache writes priced at the conservative tier; fte equivalence per section 07's stated method; the self-run's [estimated] figures name whose judgment and how.