This study tests a claim: when agents do the work, the bottleneck on a company moves from labor to attention, and the operating model becomes a routing problem for one person's judgment. The measurable quantity that follows is the attention multiplier, how much agent work one unit of human attention directs before it has to be spent again. The evidence assembled here is of three kinds. Published data from the frontier lab furthest along the curve shows attention per unit of agent work falling fast: 6.2 human turns per task down to 4.1 in six months, unsupervised agent run length up from 9.8 to 21.2 consecutive actions [v], which is roughly 5 agent actions per human touch on a task. Management science predicted the shape of the problem in 1911, 1971 and 1974, and human-factors research mapped the design space by 1978. And one modeled AI-native company opened its own operating data, which shows a multiplier of 47.6 agent actions per human message alongside a routing failure in a measurable direction: roughly two thirds of what reached its human decision queue was not a decision. The verdict is qualified. Attention is the constraint; attention-routing is a century old, and the volume it now governs is not, and the multiplier that governs it is distributed like a power law rather than a bell curve. read the post → the operating conclusion: most teams optimize for how much the ai can do; the lever is how little it can bother a human without going silent.
confidence flags throughout: [v] = primary source fetched during this study · [l] = secondary source only, primary not reachable this session · [estimated] = the authors' own judgment-involved measurement, method stated.
This study tests a claim: when agents do the work, the bottleneck on a company moves from labor to attention, and the operating model becomes a routing problem for one person's judgment. The measurable quantity that follows is the attention multiplier, how much agent work one unit of human attention directs before it has to be spent again. The evidence assembled here is of three kinds. Published data from the frontier lab furthest along the curve shows attention per unit of agent work falling fast: 6.2 human turns per task down to 4.1 in six months, unsupervised agent run length up from 9.8 to 21.2 consecutive actions [v], which is roughly 5 agent actions per human touch on a task. Management science predicted the shape of the problem in 1911, 1971 and 1974, and human-factors research mapped the design space by 1978. And one modeled AI-native company opened its own operating data, which shows a multiplier of 47.6 agent actions per human message alongside a routing failure in a measurable direction: roughly two thirds of what reached its human decision queue was not a decision. The verdict is qualified. Attention is the constraint; attention-routing is a century old, and the volume it now governs is not, and the multiplier that governs it is distributed like a power law rather than a bell curve.
When most execution is done by agents, the binding constraint on a company stops being labor and becomes the attention and judgment of the few humans left in the loop. The operating model of such a company is therefore an attention-routing system: the engineering effort goes into deciding what reaches a human, in what shape, and what never should. Token spend is not the risk; unpriced spend and un-routed attention are.
Four counter-hypotheses were written down at the same time and given equal weight throughout:
| # | counter-hypothesis | where it is answered |
|---|---|---|
| a | attention was always the constraint (Simon, 1971), so nothing here is new | §05, §10 |
| b | attention routing is management by exception rebranded | §05 (Taylor 1911, Sheridan-Verplank 1978), §10 |
| c | at enterprise scale the constraint reverts to coordination and politics, not attention | §05 (Graicunas), §09, §10 |
| d | the frontier-lab evidence is a single second-hand data point | §03, §09 |
Counter-hypothesis (a) survives the study intact and amends the claim. Counter-hypothesis (b) is half right in a way that matters. Both are treated as findings, not as objections to be defeated.
Three things are observable in this study: what companies have published about how they route work between agents and humans; what the management and human-factors literature established before agents existed; and one company's own operating record, in full, including its failures. Three things are not observable. There is no control company: a conventional twin given the same objectives, measured over the same window, is the right experiment and does not exist. There is no external audit of the self-test in §08; every one of those figures is self-administered. And there is no way to observe a human's attention directly, only its artifacts, which systematically undercounts the actions that leave no trace (reading a screen, signing into a console, a decision made and never written down).
The authors operate the company graded in §08 and hold the hypothesis being tested. The control is artifacts over impressions: every internal figure is produced by an instrument reading a system of record over a stated window with a stated denominator, re-runnable; the unflattering readings are published alongside the flattering ones and are, in this study, the more interesting half. Where a measure is self-administered and has both a floor and a ceiling, the ceiling is reported as the measure and the floor never travels alone. Where the classification involved a single human rater, the error band is published with it.
The company in §08 appears as a modeled AI-native company (~1 operator, 2 products). No internal mechanism, tool or system name appears anywhere in this report; capabilities and measured outcomes only. Spend figures are stated at published list rates or as incident magnitudes, never as actual billed amounts.
The study was prompted by a viral post claiming that an engineer at a frontier AI lab was continuously "watched" by the lab's own AI, which surfaced internal discussions and suggested what he should say. The post could not be located [l, unverified]. Two targeted searches and one direct fetch failed; only a screenshot exists. It is reported here as an unverified rumor and carries no evidentiary weight, which answers counter-hypothesis (d) directly: the anecdote is not evidence, and the study does not rest on it.
What the same lab has published is stronger and points the same way. Its December 2025 report on its own engineers [v] draws on 132 surveyed engineers, 53 qualitative interviews, and analysis of roughly 200,000 internal coding-agent transcripts, comparing February 2025 with August 2025:
| measure | Feb 2025 | Aug 2025 | population / caveat |
|---|---|---|---|
| consecutive agent actions between human check-ins | 9.8 | 21.2 (+116%) | ~200k transcripts [v] |
| human turns per task | 6.2 | 4.1 (-33%) | same [v] |
| share of own work done with the model (self-reported) | 28% (year prior) | 59% | 132 engineers, self-report [v] |
| self-reported productivity gain | +20% (earlier survey) | +50% | self-report, not measured output [v] |
| share of assisted work that would not otherwise have been done | — | 27% | self-report [v] |
The two autonomy figures are the load-bearing ones, because they are computed from transcripts rather than asked in a survey. Together they say that at the frontier, a unit of agent work consumes measurably less human attention every quarter. The self-reported figures in the same report are treated with more caution for a reason given in §05: self-reported productivity has been measured running in the opposite direction from actual productivity.
Nothing in that report, or in the same lab's ten-team case-study collection [v], corroborates the rumor's specific claims about surfacing or suggesting. No comparable primary account of internal AI use was located for any other frontier lab this session; that is a gap, not a finding.
Five public cases were reviewed. All five are [l]: the originals were posted as memos, leaks or interviews rather than as published company documents, and no clean primary was fetched.
| company | policy, dated | what happened next |
|---|---|---|
| Shopify | as reported by Forbes and Digital Commerce 360 (Apr 2025) from a circulated internal memo: "reflexive AI usage is now a baseline expectation"; teams must show AI cannot do a task before requesting headcount; no company primary fetched [l] | no reversal reported; no spend policy or budget disclosed |
| Coinbase | as reported by Fortune and PYMNTS (mid-2025) from the CEO's own podcast remarks: enterprise AI coding licenses, one week to onboard, non-compliance described as a termination matter; the CEO also stated ~40% of daily code is AI-generated; no company primary fetched [l] | the 40% figure is company-sourced with no stated denominator |
| Duolingo | as reported by TechRepublic and Entrepreneur (Apr 2025) from a LinkedIn memo: "AI-first", no contractors for automatable work [l] | per the same reporting, softened within weeks; the CEO said he had not given enough context and clarified no full-time layoffs [l] |
| Block | as reported by PYMNTS and others (2026): roughly 4,000 roles cut (about 40% of staff), which the CEO publicly framed as enabled by an internal agent platform operating asynchronously across internal tools; no company primary fetched [l] | no published spend policy located |
| Klarna | 2024: assistant handled 2.3m chats in month one, work equivalent to ~700 agents, $40m projected profit improvement [l] | May 2025 reversal: CEO stated cost had been a predominant evaluation factor and the result was lower quality; humans rehired, explicit AI-to-human handoff built [l] |
Two things follow. First, a memo is not an operating model, and the Duolingo walk-back is the cheapest available reminder that a policy citation needs a date and a check on whether it survived. Second, Klarna is the strongest public instance of the failure this study is about. The capability worked at volume. What failed was the routing: no designed path from the agent to a person, and the quality cost showed up only after fifteen months. It is the under-asking failure with a public number attached on both sides.
No company reviewed publishes a per-employee AI or token budget. That premise was searched for and not confirmed [l, gap].
| finding | value | population / window | flag |
|---|---|---|---|
| organizations using gen AI where employees review all AI output before use | 27% | McKinsey State of AI 2025 respondents whose organizations use gen AI | [l] — primary page timed out; wording consistent verbatim across multiple independent secondaries |
| experienced open-source developers using AI tools on real issues from their own repositories | 19% slower, measured | randomized controlled trial, n=16 developers, 246 issues, early-2025 models | [v] |
| the same developers' belief about the same work | 20% faster | same trial; they had forecast 24% faster beforehand | [v] |
| agentic AI projects forecast to be cancelled by end-2027 | over 40% | Gartner, June 2025 press release, cited reasons: escalating costs, unclear business value, inadequate risk controls | [l] — release 403'd; corroborated across 5+ independent secondaries |
The 19% figure carries its authors' own caveats and must not be generalized past them: it covers 16 developers on their own mature repositories with early-2025 models, and the authors themselves describe it as a snapshot they consider dated. What generalizes is the sign of the error, not its magnitude. People could not feel the attention the tools consumed. It registered as help while it was costing time, which is precisely the failure mode a self-administered attention measure is exposed to, including the one in §08.
The Gartner forecast is honest counter-evidence and is reported as such. The dominant failure mode of agentic AI projects today is not sophisticated attention-routing design. It is unclear value and cost overrun. Attention routing is the problem that bites after a project clears the basic-value bar, which per this data is a minority of projects.
Customer-service agent vendors are the one market where the price of an escalation is visible. All figures here are vendor-reported or vendor-adjacent and none was independently audited [l].
Vendor-published resolution rates (67-80% across the three vendors' own case studies, [l]) carry no stated denominator: all conversations, tickets passing a pre-filter, or resolved-without-human-touch are never distinguished. That absence is the citable finding. Independent production measurements are not published, so the gap between self-reported and measured routing cannot be sized from public data.
More useful is the pricing. At least one of these vendors is reported to charge per resolved conversation, with escalations largely carrying no charge [l, vendor-reported, unaudited]. That structure prices the ask and leaves the work uncapped, which is the commercial existence proof for the economics in §07.
If attention is the fixed input, the quantity that varies is how much agent work one unit of it directs. This report calls that the attention multiplier: agent actions directed per human touch, where a touch is a typed message, a conversational turn, or whatever unit of human engagement the measuring system can count. It is a ratio, not a productivity measure. A high multiplier says the routing produces leverage and says nothing about whether the work was worth doing; the 19% measured slowdown in §03 is the standing proof that a high ratio can coexist with a negative outcome.
Two figures in this corpus are computed from records rather than asked in a survey. They are the only anchors in the evidence gathered here.
| anchor | value | denominator · window | flag |
|---|---|---|---|
| frontier lab, per task | 21.2 consecutive agent actions against 4.1 human turns, about 5.2 : 1 | one task; ~200,000 internal coding-agent transcripts, August 2025 | [v] |
| modeled AI-native company, per message | 184,052 agent tool calls against 3,868 typed human messages, 47.6 : 1 | 30 days, 2026-07-22 to 2026-08-21, approximately one operator | [v, own records] |
They differ by an order of magnitude and they do not sit on one axis. The first is a per-task ratio across a population of engineers; the second is a per-message ratio across thirty days of one operator's whole working record, including messages that direct nothing. They bracket a range rather than trace a curve, and no arithmetic here treats them as comparable.
The largest study of individual performance located for this report is O'Boyle and Aguinis, 2012 [l]: 5 studies, 198 samples, n = 633,263 individuals across researchers, entertainers, politicians and athletes. Individual performance is not Gaussian. It follows a Paretian distribution, consistently across industry, job type, performance measure and time frame, with a long right tail of disproportionate contributors. That finding is about human output in general. It licenses drawing the multiplier's distribution as a power law rather than a bell curve, and it is not evidence that AI leverage is itself distributed that way.
The pre-agent spread between engineers is smaller than folklore says. The famous 28:1 comes from Sackman, Erikson and Grant, 1968 [l], where it is the debug-time extreme from n = 12 programmers on two tasks. Dickey, 1981 [l] traced it to a single pairing confounded by machine language and batch turnaround against a time-shared high-level language, and attributed roughly 5:1 to the programmer. Modern consensus lands near 10x best to worst [l], and 10x is the number this report uses. Contribution concentration runs the same direction [l]: one in four active repositories is single-authored, only about 9.6% have two or more long-term collaborators, and roughly 40% of commits come from occasional contributors submitting between one and ten changes in total.
No per-user token distribution is published for any AI coding tool as of 2026-08-22. That is the honest headline of this section. What exists is adjacent and does not substitute: usage bands from one vendor, a stated two-to-threefold spread between a typical daily user and a power user rather than a percentile distribution; concentration data from another that is geographic and task-level rather than per person; per-user cost figures circulating on aggregator blogs with no primary behind them, not cited here as fact. Roughly 70% of engineers use two to four AI tools at once [l], so no single vendor's data could produce an industry-wide per-person distribution even if one were released tomorrow. Any quantile ladder is therefore an estimate, and it is labelled as one below.
Units: agent actions directed per human touch. Every row except the two marked measured is an estimate. The shape rests on O'Boyle and Aguinis for the power-law form and on the two anchors for scale. This table is an illustrative construction, not a finding.
| quantile | est. multiplier | who is plausibly here | basis | type |
|---|---|---|---|---|
| top 0.1% | 50-100:1 | frontier-lab engineers on agent-native workflows; solo agent-native operators | self-test measured 47.6:1 at the low end of this band; a >80%-merged-code environment implies at least this | estimate (one measured anchor) |
| — | 47.6:1 | the modeled AI-native company (self-test) | 3,868 human messages against 184,052 agent tool calls, 30-day window | measured |
| top 1% | 15-40:1 | frontier-lab engineering broadly; agentic-IDE power users | interpolation between the 5.2:1 task-level anchor and the self-test | estimate |
| — | ~5.2:1 | frontier-lab engineers, task-level, Aug 2025 | 21.2 tool calls / 4.1 human turns per task | measured |
| top 10% | 5-15:1 | heavy agentic-IDE users; the 34% of organizations claiming >60% AI code | the Aug-2025 task ratio as the floor; a 2h/day median as the ceiling constraint | estimate |
| median (50%) | 1-3:1 | the 90% of developers using AI daily at about 2h/day; the measured 22-27% AI-authored share | measured share implies assistance, not delegation | estimate |
| bottom 10% | ~0-1:1 | chat-only usage; the roughly two thirds of enterprises not scaling agents; the RCT population (measured negative: 19% slower) | ≤10% scaling agents; the RCT | estimate, with one measured negative |
Three caveats travel with this table. (1) The multiplier is a ratio, not a productivity measure; the RCT is the standing proof that a high ratio can coexist with a negative outcome. (2) The 80%-reliability horizon of about 27 minutes caps how far the ratio can rise before quality degrades, which is why the ladder's top is bounded by routing design rather than by model capability. (3) The two measured rows come from two different denominators, per-task against per-message over 30 days, and are not directly comparable; they bracket the range, they do not sit on one axis.
| company | stated AI intensity | scope stated | date | flag |
|---|---|---|---|---|
| Anthropic | >80% of code merged into the production codebase authored by its model; 90%+ including scripts and experimental code, stated as a leadership estimate | merged code | May 2026 | [v] |
| OpenAI | 97.9% of employees using its coding agent; 99.8% of employee-generated output tokens through it | internal usage, company cautions it is not representative of customers | Jun 2026 | [l] |
| "well over 30%" of code checked in generated with AI assistance | new code checked in, all engineering | Apr 2025 | [l] | |
| Microsoft | "maybe 20%, 30%" of the code inside its repos written by software | repos, scope unspecified; more in Python, less in C++ | Apr 2025 | [l] |
| Coinbase | about 40% of daily code AI-generated | daily code, denominator never defined | Sep 2025 | [l] |
| Block | roughly 4,000 roles cut, framed by its CEO as enabled by an internal agent platform; engineers said to have shipped 40% more code | headcount and code volume | Feb 2026 | [l]; contested: a named sell-side analyst told the Wall Street Journal that the vast majority of the cuts were probably not due to AI |
| typical large enterprise | about 6% are AI high performers by EBIT impact; at most 10% scaling agents in any single function | survey population, n and date unconfirmed | 2025-26 | [l] |
Only the first row is measured against a published denominator. Every other intensity figure is self-reported by an executive, and "code checked in," "code in our repos" and "daily code" are three different things, none specifying lines, characters, commits or accepted completions. Read them against the measured discount: across 425 instrumented organizations, AI-authored code share was 22% in late 2025 and past 27% in early 2026, while 34% of organizations self-reported more than 60% [l]. That is a two-to-threefold gap between what companies say and what their repositories show, the same failure shape as the perception gap in §03 one level up.
The multiplier is bounded by how long an agent runs before a human re-engages, and the closest published proxy is a time horizon: the human-task duration at which a model succeeds some fraction of the time, fitted against paid expert baseliners [v]. The revision published on 2026-01-29 [v] puts the 50% horizon for the strongest model measured at about 5 hours, doubling every 196 days overall and every 89 days when fitted from 2024 onward. The 80% horizon governs routing and is far shorter: about 27 minutes for that same model, against a 50% horizon of roughly 4 hours 49 minutes whose confidence interval runs from under two hours to over twenty [v]. The 80% figure has stayed in the tens of minutes release over release while the 50% figure has climbed into hours.
That gap is the whole argument. Capability sets what an agent could finish; reliability sets how often a person must be touched, and reliability is moving slowly. The multiplier is therefore an engineering variable set by routing design rather than a property of the model, which is why it can be raised this quarter without waiting for the next release.
Simon, 1971 [v, quote verified against the original this session]:
"What information consumes is rather obvious: it consumes the attention of its recipients. Hence a wealth of information creates a poverty of attention and a need to allocate that attention efficiently among the overabundance of information sources that might consume it."
Simon's claim is that attention scarcity is a general property of any information-processing system, not an artifact of human labor. This is counter-hypothesis (a), and it holds. What changes when execution is agentic is the volume competing for that attention: it scales with compute rather than with headcount.
Galbraith, 1974 [l] framed the organization as an information processor whose design problem is the gap between processing capacity and processing load, and named four strategies. Each maps onto a routing choice available to an agent-heavy company (the mapping is this study's, not Galbraith's):
| Galbraith's strategy (1974) | the agent-routing equivalent |
|---|---|
| create slack resources | widen the band in which an agent decides without asking, trading precision for less escalation traffic |
| create self-contained tasks | give a role full authority and resources over a bounded domain so routine cases never escalate |
| invest in vertical information systems | the inform tier: a durable record a human can sample without being interrupted |
| create lateral relations | agent-to-agent resolution of cross-cutting issues, so nothing routes up |
An agent-heavy company has a fifth lever Galbraith did not: raising raw processing capacity per unit of headcount. That changes the marginal cost of the first three without removing the need for any of them.
Taylor, 1911 [l] named the exception principle: the manager receives only condensed, summarized reports showing deviations from standard, so attention can go to the significant ones. This is counter-hypothesis (b), and it is half right. The taxonomy in §06 is management by exception. What is new is the ratio it operates on.
Graicunas, 1933 [l] gave the combinatorics of that ratio: R = n(2^(n-1) + n - 1) relationships for n subordinates, giving 44 at n=4 and 1,080 at n=8. The premise is Simon's: the manager's attention is fixed, and coordination combinatorics, not headcount, exhausts it. This is also the strongest support for counter-hypothesis (c), and it cuts both ways. Combinatorial coordination load is what returns at enterprise scale, and it is exactly what a one-operator company does not yet face.
Mission command [l] pushes decision rights to the lowest competent level and gives intent plus constraints rather than method, explicitly to conserve the senior commander's attention for decisions only they can take. It is the decide tier pushed to the edge, with a named re-sync point instead of per-step permission.
Interruption cost [l]: the widely cited figure is 23 minutes and 15 seconds as the average time to return to an original task after an interruption, from field observation of knowledge workers. The caveat has to travel with it: over 80% of interrupted work was resumed the same day, and workers typically completed one or two other tasks before returning, so the figure conflates task-switching cost with task-queuing behavior and overstates pure cognitive resumption.
Decision fatigue [l] is presented here as disputed, not settled. The 2011 parole-board analysis of 1,112 rulings reported favorable rulings falling from roughly 65% after a break toward zero before the next one. A same-volume comment and a 2016 methodological critique argue the pattern is confounded by case ordering, since routine and unrepresented cases are systematically scheduled later. It is one of the most-cited pieces of decision-fatigue evidence in the popular press and one of the most contested in the literature.
Bainbridge, 1983 [v via a summary confirming the argument; primary paywalled] states the irony this whole class of system inherits: automating the routine parts of a task leaves the human responsible for exactly the rare, poorly specified, high-stakes cases, while removing the routine practice that kept them skilled at it.
Parasuraman and Riley, 1997 [l] name the four failure modes of the human-automation relationship: use, misuse (over-reliance beyond validated competence), disuse (ignoring automation that works, usually after a false alarm), and abuse (automation imposed without regard for the operator's actual task).
Sheridan and Verplank, 1978 [l] give the ten-level automation ladder whose midpoints are precisely the design space in §06: the computer suggests one alternative (4), executes on approval (5), allows a limited veto window before acting (6), executes and necessarily informs (7). The three shapes below map onto levels 4-6 (escalate), 7 (inform) and 7-8 (decide silently). The design space was fully described for undersea teleoperation forty-plus years before LLM agents existed, which is the honest version of counter-hypothesis (b).
Everything an agent does arrives at a human in one of three shapes. The taxonomy is a century old (§05). It now governs a stream of tens of thousands of actions a month rather than a handful of decisions a day.
| shape | cost to the human | failure if over-used | failure if under-used | evidence |
|---|---|---|---|---|
| decide silently — agent acts, says nothing | zero | actions with money, credential, deletion or outbound consequences happen unseen; the human's model of system state goes stale, and Bainbridge's skill decay follows | every routine action becomes an interruption; the human is the throughput ceiling | Klarna: full routing to AI with no designed handoff produced a quality failure the company publicly reversed [l] |
| inform — agent acts, leaves a durable record | a glance, deferrable | volume defeats it; a record nobody reads is indistinguishable from silence | consequential actions become invisible after the fact and cannot be audited | Taylor's exception principle: condensed reports of deviations, not the stream [l]; the vendor-tier information system in Galbraith [l] |
| escalate — agent stops and asks | a context switch, a judgment, a reply, plus the agent idling throughout | the human becomes the bottleneck and cannot feel it happening | one-way doors get taken without a human; the never-silent floor is breached | METR: 19% measured slowdown against a 20% believed speedup, n=16, early-2025 models [v] |
Two properties of this table matter more than the taxonomy.
First, the costs are wildly asymmetric and only one of them is visible. Silence costs nothing and shows nothing. An escalation costs a context switch measured in the tens of minutes [l] plus the agent's idle time, and shows up as a message that looks like diligence. A system whose default is to ask will look careful and behave like a bottleneck, and the METR perception gap says the people inside it will not notice.
Second, the two failure directions are not symmetric in when they are discovered. Over-asking is discovered never, because every individual ask is defensible. Under-asking is discovered late and publicly, as Klarna's fifteen months show. That asymmetry is an argument for measuring the routing rather than trusting it, which is what §08 attempts.
If attention is the scarce input, the useful unit is cost per decision that reaches a human, not cost per token.
The cost of an ask. Take the interruption figure at face value and then bracket it. A human fielding 50 asks a day, each carrying a full context switch, spends time as follows:
| recovery cost per ask | asks/day | human hours/day consumed | assumption |
|---|---|---|---|
| 23 min | 50 | 19.2 | the field-observation figure taken literally [l] |
| 10 min | 50 | 8.3 | bearish: strips out the task-queuing conflation the critique identifies |
| 10 min | 20 | 3.3 | bearish cost, a filtered queue |
| 23 min | 20 | 7.7 | literal cost, a filtered queue |
The assumptions are exposed because they carry the conclusion. Even at the bearish 10 minutes, a 50-ask day is a full working day before a single decision is made, and the queue that produces it can be generated by agents in minutes. Both rows in the 50-ask case are arithmetic on a contested input, not a measurement; what survives the bracketing is the shape. The number of asks, not the cost of any one of them, is the lever.
The mirror failure: capping instead of metering. The instinctive control on agent spend is a hard limit. It looks like prudence and behaves like an outage. Two incidents from the company in §08 make the point in opposite directions:
| incident | what happened | what it cost |
|---|---|---|
| free-tier allowance, 2026-08-04 | 2,947 of 3,000 monthly CI minutes consumed by day 4 of a 31-day cycle (98% of cap, ~7.6x a linear burn), discovered only via the vendor's own 90% email; every internal cost report had correctly shown $0.00 throughout, because the meter read dollars and the resource was free | four days of unmonitored ramp; a truthful zero carrying no information |
| vendor spending limit, 2026-08-16 | a shared paid-API account hit a vendor-side spending limit and returned errors to every caller across two products for several hours; the account's own spend for the period was approximately $3 | a capacity outage across two products, caused by the control rather than by the cost |
A capped agent and an over-asking agent fail identically: both hand the work back to the human, and both do it exactly when the system is busiest. The instrument that survives is a meter plus an alert on the projected total, with no switch. The customer-service vendor charging per resolution and nothing for an escalation [l] is the commercial version of the same design: it prices the ask and leaves the work uncapped.
Attention as denominator, stated. Three ratios are worth computing in any agent-heavy operation and none of them involves tokens: agent actions per human message, human messages per completed work item, and the share of items reaching the human that needed to. The first two say how much leverage the routing produces. The third says whether it is any good. §08 reports all three.
Subject: a modeled AI-native company, approximately one operator and two consumer products, all execution performed by agents. Assessment window 2026-07-22 to 2026-08-21 (30 days) unless a row states otherwise. Every figure is self-administered. Where a measure has a floor and a ceiling, both are reported and the ceiling is the measure.
| measure | value | window · denominator | flag |
|---|---|---|---|
| agent tool calls per human typed message | 47.6 : 1 | 184,052 tool calls against 3,868 typed messages, 30 days | [v, own records] |
| human typed messages per day | 138 mean, 125 median, 366 peak | same window | [v] |
| human messages per completed work item | 2.18 | same window | [v] |
| work items with a confirmed human execution step | 1.7% confirmed · 8.3% ceiling | stratified sample of 60 completed items; unrecorded out-of-system steps counted against | [v, ceiling is the measure] |
| code changes in a sampled stratum written by a human | 0% measured, 13% 95% upper bound | 25 sampled non-trailer commits | [v, floor never travels alone] |
One full audit of the human decision queue, 75 items, each answered in full by the human and then re-scored against whether the answer was needed:
| stage | count | share of 75 |
|---|---|---|
| items routed to the human's decision queue | 75 | 100% |
| not a decision at all | 37 | 49% |
| survived the initial filter | 38 | 51% |
| of those 38, human explicitly said an answer was not needed | 16 | 21% of 75 (42% of the 38) |
| genuinely needed the human's judgment | 22 | 29% |
The filter was wrong on 53 of 75 items, roughly two thirds. Half the queue was not a decision; of the half that looked like one, another 42% still resolved without the human. The waste is upstream of the human ever reading the item, which makes it routing-shaped rather than judgment-shaped. Three standing exemptions were carved directly out of this measurement and are now never asked again: instrumentation additions, error investigations, and any choice with only one viable branch. The source does not decompose the 53 by category, so no further split is claimed [flagged].
Upstream of that queue, on one product's work tracker over a 12-day window: 1,256 items filed, 662 closed (53%), net queue growth positive on every single day; 242 of the filed items (~19%) were records of completed work with nothing left to do; of 241 items filed at the highest priority, 54 (22%) were genuinely urgent on review [v, own records].
A rolling scan of the company's own outbound messages, classified by shape and scored against a per-category length target. 2,256 messages classified. Window caveat: the instrument's reported window is imprecise; it is most likely the trailing 14 days and is not independently confirmed to be 30 [flagged].
| message type | n | median words | over target | on target | verdict |
|---|---|---|---|---|---|
| decision-ask | 233 | 515 | 101 (43%) | 95 (41%) | too long |
| incident | 237 | 451 | 103 (43%) | 89 (38%) | too long |
| status | 223 | 399 | 85 (38%) | 97 (43%) | too long |
| close-out | 209 | 355 | 55 (26%) | 94 (45%) | trending too short |
| unclassified | 1,354 | 335 | 332 (25%) | 774 (57%) | too long |
On-target never exceeds 57% in any category. The two highest-stakes shapes by design, the decision-ask and the incident report, are also the two most likely to run long, at 43% over target each. That is the exact inverse of what a scarce-attention system should produce.
| measure | value | denominator · caveat | flag |
|---|---|---|---|
| standing behavioral rules whose function is routing attention | 19 of 68 (28%) | full rulebook; single-rater manual classification, a second rater could move 3-4 rules, so ±4 points | [estimated] |
| items in a human-review workflow state on the larger product | 5 of 3,887 created in the window (0.13%) | current snapshot only; no state-history query run | [v] |
| outages caused by a hard spend cap | 1, across 2 products, over ~$3 of spend | single incident, 2026-08-16; no base rate available | [v] |
| free-tier allowance consumed unnoticed | 98% of cap by day 4 of 31 | single incident, 2026-08-04 | [v] |
| approved budget overrun by per-run rather than cumulative tracking | approved ceiling exceeded roughly 4-5x by tracking each run's marginal cost instead of the cumulative total | single incident, 2026-06-15 | [estimated] |
| top-tier model capacity spent on mechanical rather than judgment work | 71% of 25% of weighted burn | ~7-week audit window | [v] |
| stated first-person commitments made and not executed | 1 recorded instance | ledger one day old at measurement; no base rate establishable | [flagged, n=1] |
| throughput change | 3.0x (89 → 268 completed items/week) | 4-week average over a 17-week window, previously published | [v, self-administered] |
| defect rate change | 2.4x lower (131 → 54 filed bugs per 100 completed items) | same window; rose again on a second product launch | [v, self-administered] |
| output-scale equivalence | roughly 35-40 people's full-time output, one director | previously published, judgment-involved, self-selected population | [estimated] |
One. The leverage is real and the routing is not good. One human directs roughly 48 agent actions per typed message and fewer than one work item in twelve requires a hand at the ceiling reading. In the same operation, seven in ten items that reached the decision queue should not have. Both are true, and the second is the largest single recoverable cost in the system.
Two. The instruments disagree about how much attention costs, and the disagreement is unresolved. The message instrument says the asks that do arrive run long 43% of the time. The queue audit says most of them should have arrived as information rather than as a question at all. Fixing the second makes the first worse in percentage terms while making it better in absolute terms, and no measure in this corpus reconciles them.
Three. What gets measured here is what leaves an artifact. The human's least artifact-producing actions, reading a screen, signing into a console, a decision taken aloud and never written down, are exactly what the record misses. That is why the 1.7% confirmed figure travels with an 8.3% ceiling and why the 0% commit reading travels with a 13% upper bound. Both readings say the same thing: the instrumentation cannot yet see what the human executes.
n = 1, and it is the author's. One company, self-administered, measured by instruments its own operators built to test a hypothesis they hold. The controls in §02 reduce the risk that a number is wrong; they do nothing about the risk that the wrong numbers were chosen.
Scale reversion is untested and probably real. Every internal figure comes from an operation with approximately one human. Graicunas's combinatorics [l] say that coordination load grows non-linearly with the number of humans, which means the constraint identified here may be an artifact of having exactly one person to route to. Counter-hypothesis (c) is not answered by this study and cannot be answered by it. A two-person or ten-person agent-heavy company is a different measurement.
The frontier-lab evidence is self-reported in the places that flatter it. The transcript-derived autonomy figures (9.8 → 21.2, 6.2 → 4.1) are computed from records [v]. The productivity figures in the same report (+50%, 59% of work) are surveys of the people using the tool, and the one rigorous randomized measurement available found self-reported speed running 39 points opposite to measured speed [v]. Those two classes of figure should never be cited as if they were the same kind of evidence.
Survivorship. The public companies with published routing designs are the ones that survived long enough to publish. Gartner's forecast that over 40% of agentic AI projects will be cancelled by end-2027 [l] describes a population this study cannot see, and the reasons given (cost, unclear value, weak risk controls) are not the failure this report is about.
The interruption figure is contested and load-bearing. §07's model rests on it, which is why it is bracketed at 10 and 23 minutes rather than asserted. The decision-fatigue evidence is weaker still and is presented as disputed.
| claim | verdict | confidence |
|---|---|---|
| in an agent-heavy company, the binding constraint is human attention rather than labor | supported | strong inference — direct measurement at n=1 (47.6:1, 2.18 messages per item), consistent with the frontier lab's falling human turns per task [v] |
| attention scarcity is new | rejected | verified — Simon stated it in 1971 [v]; counter-hypothesis (a) holds |
| the routing taxonomy is new | rejected | verified — Taylor 1911, Sheridan-Verplank 1978 [l]; counter-hypothesis (b) holds |
| what changed is the ratio of actions to decisions | supported | strong inference — 47.6 agent actions per human message measured internally [v]; 21.2 consecutive actions per check-in at the frontier [v] |
| the attention multiplier is power-law distributed and routing-bound | supported | likely — the power-law form is borrowed from a 633,263-person study of human output generally [l], not measured on AI leverage; no per-user token distribution is published for any AI coding tool, so the ladder in §04 is an estimate against two measured anchors an order of magnitude apart. The routing-bound half is the firmer one: an 80%-reliability horizon of about 27 minutes against a 50% horizon of hours [v] means touch frequency is set by routing design rather than by model capability |
| companies fail in both routing directions, and over-asking is the invisible one | supported | strong inference — Klarna for under-asking [l]; METR's 39-point perception gap for the invisibility of over-asking [v] |
| a hard spend cap is a worse instrument than a meter and an alert | supported | verified at n=1 — one cap produced a two-product outage over ~$3 of spend; likely as a general claim, single incident, no base rate |
| an agent-heavy operating model is substantially an attention-routing system | supported | likely — 28% (±4) of one company's standing rulebook governs what reaches its human, single-rater classification, no external baseline exists to compare against |
| the constraint holds at enterprise scale | untested | Graicunas predicts coordination reverts to the constraint [l]; this study cannot see it |
The hypothesis is sharpened, not confirmed as written. Attention was always the constraint and the routing taxonomy is a century old. What is new is the volume the taxonomy now governs and the asymmetry of its failure modes: over-asking is defensible item by item, invisible to the person paying for it, and compounding; under-asking is discovered late and in public. The engineering effort in an agent-heavy company therefore goes into measuring a century-old routing model at a volume where nobody can feel whether it is working. That is the claim this study would defend.
Review date: 2026-11-21, or when the first external company publishes a queue-composition measure against this framing, whichever comes first.
| source | what it supplied |
|---|---|
| Anthropic, "How AI Is Transforming Work at Anthropic" (2025-12-02) — anthropic.com/research/how-ai-is-transforming-work-at-anthropic | 132 engineers surveyed, 53 interviews, ~200k transcripts; 9.8 → 21.2 consecutive actions; 6.2 → 4.1 human turns per task; 28% → 59% share of work; +20% → +50% self-reported gain; 27% net-new work |
| Anthropic, "How Anthropic Teams Use Claude Code" (2025-07-24) — claude.com/blog/how-anthropic-teams-use-claude-code | ten-team case studies; no corroboration of the viral claim |
| Anthropic Economic Index, June 2026 — anthropic.com/research/economic-index-june-2026-report | external-user delegation and engagement figures; explicitly does not publish a human-review share |
| METR, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity" (2025-07-10) — metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ | RCT, n=16, 246 issues: 19% measured slowdown against a 20% believed speedup, with the authors' scope and staleness caveats |
| Simon, "Designing Organizations for an Information-Rich World" (1971), in Greenberger ed., Johns Hopkins Press | the poverty-of-attention passage, verified against the original |
| Bainbridge, "Ironies of Automation," Automatica 19(6), 1983 (argument confirmed via summary; primary paywalled) | the two ironies of automation oversight |
| Anthropic, "When AI Builds Itself" (June 2026) — anthropic.com/institute/recursive-self-improvement | >80% of merged code authored by the model (May 2026), up from low single digits in Feb 2025; the 90%+ figure flagged as a leadership estimate |
| METR, "Measuring AI Ability to Complete Long Tasks" (2025-03-19) — metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/ | the time-horizon definition and method, fitted against paid expert baseliners |
| METR, "Time Horizon 1.1" (2026-01-29) — metr.org/blog/2026-1-29-time-horizon-1-1/ | revised 50% horizons on a 228-task suite; ~5h for the strongest model measured; doubling 196 days overall, 89 days from 2024; the ~27 min 80% horizon against a 4h49m 50% horizon |
| the modeled company's own operating records, ledgers and instruments | every figure in §08 |
| source | what it supplied | why it is tier 2 |
|---|---|---|
| McKinsey, State of AI 2025 — mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai | 27% of organizations using gen AI review all output before use | primary page timed out; wording consistent verbatim across multiple independent secondaries |
| Gartner press release, 2025-06-25 | over 40% of agentic AI projects cancelled by end-2027; Jan 2025 poll of 3,412 attendees | release returned 403; corroborated across 5+ independent secondaries |
| Klarna reversal coverage (2024 announcement; May 2025 CEO statements) | 2.3m chats in month one, ~700 agents equivalent, $40m projected; "cost was a predominant evaluation factor", "lower quality", rehiring and explicit handoff | five independent outlets, consistent figures; no Klarna-authored primary fetched |
| Shopify (Apr 2025), Coinbase (mid-2025), Duolingo (Apr 2025 + walk-back), Block (2026) | the mandate and reversal table in §03 | memos circulated as leaks or interviews; no clean company primary |
| Galbraith, Interfaces 4(3), 1974 | the four information-processing strategies | primary not fetched; taxonomy consistent across org-design literature |
| Taylor, Shop Management (1911) | the exception principle | 1911 text not fetched; wording is a source-consistent reconstruction, not scan-verified |
| Graicunas, "Relationship in Organization" (1933) | R = n(2^(n-1) + n - 1); 44 at n=4, 1,080 at n=8 | original not freely available; formula sanity-checked arithmetically |
| US Army ADP 6-0, mission command doctrine | intent plus constraints, decision rights at the lowest competent level | doctrine PDFs not web-fetchable |
| Mark, Gudith & Klocke, CHI 2008, and the field studies behind it | 23 min 15 s average return-to-task, with the task-queuing caveat | ACM paywalled; caveat traced through a critical secondary analysis |
| Danziger, Levav & Avnaim-Pesso, PNAS 108(17), 2011; Weinshall-Margel & Shapard comment; Glöckner (2016) | the parole-board finding and its case-ordering critique | neither primary fetched; presented as disputed |
| Parasuraman & Riley, Human Factors 39(2), 1997; Sheridan & Verplank (MIT, 1978) | use/misuse/disuse/abuse; the ten automation levels | canonical, neither fetched; no verbatim quotes used |
| Intercom, Sierra, Decagon resolution rates and pricing | the vendor marketing-vs-production gap; outcome pricing that charges nothing for an escalation | vendor-reported or vendor-adjacent, unaudited, no stated denominators |
| O'Boyle & Aguinis, Personnel Psychology 65:79-119 (2012) | 5 studies, 198 samples, n=633,263; individual performance is Paretian rather than Gaussian | journal paywalled; figures and finding consistent across independent summaries |
| Sackman, Erikson & Grant, CACM 11:1 (1968); Dickey (1981); Prechelt TR 1999-18 | the 28:1 claim, its debug-time confound, and the ~10x modern consensus | none of the three primaries fetched; the correction is consistent across all secondary treatments |
| mining-software-repositories literature (arXiv 2308.08991, arXiv 1407.2535) | one in four active repositories single-authored; ~9.6% with 2+ long-term collaborators; ~40% of commits from occasional contributors | abstracts only, not fetched to full text |
| DORA 2026 | 425 instrumented organizations: AI-authored code 22% late 2025, past 27% early 2026, against 34% of organizations self-reporting >60%; 90% of developers using AI, median 2h/day | reported via secondary; the instrumented population is stated but the primary was not fetched |
| OpenAI internal coding-agent figures (2026-06-11) | 97.9% of employees on the coding agent; 99.8% of employee-generated output tokens through it | reached this study via secondary reporting of OpenAI's own release; primary not fetched |
| Google (Q1 2025 earnings call), Microsoft (LlamaCon, Apr 2025), Coinbase (Sep 2025) | executive "share of our code is AI" claims: >30%, 20-30%, ~40% | all executive statements with no published denominator; read against the DORA discount |
| Block AI attribution, incl. Wall Street Journal reporting of a named sell-side analyst's dissent | the 40%-more-code claim and the contested link between the agent platform and the cuts | company framing plus one on-record analyst rebuttal; no independent measurement exists |
| McKinsey State of AI | ~6% AI high performers by EBIT impact; at most 10% scaling agents in any single function | via secondary; n and publication date unconfirmed |
| JetBrains developer survey 2026 (~10,000 developers) | 70% of engineers use 2-4 AI tools simultaneously | via secondary; supports the claim that no single vendor could publish an industry-wide per-person distribution |
| vendor usage bands and per-user cost figures for agentic coding tools | a stated 2-3x typical-to-power-user spread; aggregator per-user cost figures | usage bands are vendor marketing; the cost figures appear only on aggregator blogs with no primary and are not cited as fact |
| the viral post that prompted the study | nothing | could not be located; a screenshot only. Reported as unverified rumor, carries no evidentiary weight |
open verification debt before wider distribution. The McKinsey 27% figure and the Gartner cancellation forecast are the two load-bearing external numbers that were not opened at their own primaries; both should be re-fetched before publication. The Shopify, Coinbase, Duolingo and Block memos are the most-quoted items in §03 and are the least well-sourced. No published per-employee AI or token budget was located at any company, and the report asserts none. Stanford HAI's AI Index was not fetched this session and supplies nothing here.