{"path":"research/the-agent-population-and-what-checking-is-up-against.md","content":"# How many agents, and what checking is up against: Epoch's estimate read into the corpus (2026-10-06)\n\n**Status**: research, nothing ruled. The proposals are ledger rows U14–U17 and X11\n(`session26-open-rulings.md`). **Origin**: the founder, sending Epoch AI's report on 2026-10-06: *\"This is\nalso relevant for how Deliberus could run with much more AI features woven in but also is relevant in\nterms of 'what we're up against' both as it relates to alignment and [generation/persuasion vs vetting and\nsensemaking capacity] etc\"*.\n\n**Read depth**: Jason Li, *How many AI agents could run on the AI chips shipped through 2027?*, Epoch AI,\n2 October 2026, read in full with its appendices and notes. Its code and trace data were not run. Every\nfigure below attributed to the report is the report's; the arithmetic that combines them with this\ncorpus's own figures is marked as ours.\n\n---\n\n## 0. In short\n\n- **The generation side now has a number.** Memory shipped through 2027 could run about 30–170 million\n  frontier-model agents at once, or around two billion on a cheaper open model, if all of it were\n  deployed and given to that work. Running all week, the frontier agents would put in as many hours as\n  140–720 million full-time employees. That is a ceiling on capacity, not a forecast of use, and the\n  report itself warns that demand may lag.\n- **Volume is the wrong unit, twice over.** For machine work we run, what compounds is the depth of\n  machine building on machine, not the amount (ruled 2026-09-29). For agreement among agents, what counts\n  is how many independent model families stand behind it, not how many instances: a million agreeing\n  agents may be a handful of samples.\n- **The agents are on both sides.** The same fleet that can flood a reader can check for them, since every\n  reader can have an assistant. Persuasion needs only compute. Checking needs compute *and a record to\n  check against*, or the reader's assistant is one more persuadable judge. That record is the slot this\n  project claims.\n- **Against a flood, checking keeps pace only by reuse, so the race is decided where sameness is\n  decided.** A check made once answers every paraphrase only if the paraphrase is recognised as the same\n  claim, and paraphrase costs nothing at this scale.\n- **For weaving more machine work into Deliberus: single questions, not agent sessions, and measurement\n  first.** A frontier agent session costs $15–50 an hour in the report's traces. The keeper's daily cap\n  buys about four minutes of one at the $30 reference, where a whole extraction cost 19 cents per thousand words as single\n  calls. With the price of a given capability falling fast, the scarce input in a year is a pass proven to the bound its use needs,\n  not compute.\n\n## 1. What the report measures, and what it does not\n\n**The method.** High-bandwidth memory (HBM) is the shared bottleneck of AI accelerators, made by three\ncompanies (Micron, Samsung, SK hynix). The report counts HBM shipped from 2025 on, converts it to units of\none GB300 GPU, and multiplies by how many agent sessions one such unit can serve:\n- **Open models**: measured directly in a serving benchmark that replays real coding-agent traces\n  (SemiAnalysis's AgentX). DeepSeek V4 Pro serves 31.4 sessions per GPU at 50 output tokens a second per\n  user, 14.4 at 100.\n- **Closed frontier models**: inferred from what a continuously running agent costs at API prices, about\n  $30 an agent-hour as the reference, and an assumed ratio of 5–10 between API revenue and serving cost.\n  That gives 0.83–1.67 sessions per GPU.\n\n**The figures.**\n\n| | Through 2026 | Through 2027 |\n|---|---|---|\n| Concurrent frontier agents, central hardware assumption | 20–40 million | 50–101 million |\n| The same, with the hardware assumption varied | 16–56 million | 33–171 million |\n| Full-time-equivalent working hours | 67–240 million people | 140–720 million people |\n| DeepSeek V4 Pro at 50 tokens/s, same hardware | 0.6–1.1 billion | 1.2–3.2 billion |\n| API-equivalent spending at 20% effective use | $1.1–2.1 trillion a year | $2.6–5.3 trillion a year |\n\nFor scale, the report gives the United States 342 million people and about 100 million knowledge workers,\nand model developers' combined revenue at over $100 billion a year, reaching about $1 trillion by the end\nof 2027 if it keeps growing fivefold a year.\n\n**What an hour of an agent costs today**, from logged Codex and Claude Code sessions with waits for a\nhuman removed: $18 for GPT-5.5, $16 for GPT-5.6 Sol, $24 for Claude Opus 4.8, $50 for Claude Fable 5, and\nmore for sessions with very long contexts ($39 and $105).\n\n**Limits the report states itself, which travel with every figure.**\n- It counts **capacity under full deployment and allocation**, not use. Chips can sit uninstalled for\n  want of powered data-centre space, and *\"These estimates suggest a risk that the compute buildout could\n  run ahead of inference demand.\"*\n- An \"agent\" is **a session inside a harness such as Claude Code**, model calls and tool use together. The\n  workload measured is coding. *\"Agents can also produce output much faster than humans, though the\n  quality of that output varies.\"*\n- The two main assumptions move the totals severalfold: the revenue-to-cost ratio, and how much more\n  newer memory serves per gigabyte.\n- **Price per capability is falling**: the report cites Epoch's estimate of a 47% decline per quarter\n  since 2023 across the benchmarks studied, and notes that this cuts both ways, while cheaper tasks may\n  raise total demand (Jevons).\n\n## 2. What checking is up against\n\n### 2a. The ratio has a number\n\nThe scrutiny gap's sentence was *\"The volume and fluency of claims is outrunning the capacity to check\nthem\"* (`the-scrutiny-gap.md` § 1), with information throughput as the measured mechanism of machine\npersuasion (§ 2). The report sizes the volume side. Our arithmetic: hardware shipping now could supply\n1.4 to 7 times the working hours of every knowledge worker in the United States, and the people who check\ndo not multiply. One percent of the central 2027 estimate is half a million to a million concurrent\nfrontier agents. An actor who wants to persuade needs a sliver of the fleet, not most of it.\n\nTwo qualifications keep this honest. Most of the fleet will be paid to do other work. And the report's\nagents are coding agents, so their output rate says nothing directly about persuasion. What the number\nfixes is the order of magnitude a defence must be designed against: a ratio of machine output to human\nattention that moves from small to large on hardware already being built.\n\n### 2b. The population is smaller than it looks\n\nThe corpus already holds that two instances of one model are one detector sampled twice\n(`the-scrutiny-gap.md` § 5; the daemon co-stimulation rule needs an independence condition), that agent\npopulations converge when they read the same public board (§ 6), and that models agree on wrong answers\nmore as they get more accurate (Correlated Errors, ICML 2025). The assembly-theory lens names the\ndistinction: agreement is either **copy number** (shared lineage) or **convergence** (independent lineages\npushed together by the world), and only the second is evidence\n(`why-representations-converge-and-whether-minds-do.md`).\n\nSo **a population of a hundred million agents may be little more than a handful of model families,\nsampled many times.** Two\nconsequences:\n- **Fabricated consensus becomes nearly free.** *\"Three experts disagree with you\"* was the strongest\n  pressure tactic against model judges, flipping 60.3% of the verdicts of a model otherwise stable under\n  pressure (`steadying-an-outside-judge.md`). A\n  fleet can make that sentence literally true by posting the view under many names. The lookup that\n  defeats fabricated consensus works only where **attribution in the record is harder to fabricate than\n  the consensus**: published authors, extracted with provenance, as the graph holds today.\n- **The sybil hole is priced.** The threat model says no layer has a sybil or source-independence defence.\n  Read beside this report, that is not a hole a determined actor would find costly to use.\n\n### 2c. The agents are on both sides\n\nThe fleet that can flood a reader can also check for them: every reader can have an assistant, and those\nassistants draw on the same capacity the report counts. **The asymmetry is not compute. Persuasion needs only\ncompute; checking needs compute and a record to check against.** Without a record, the reader's assistant\njudges in the conversation, where verdicts flip under pressure (25–71% scripted, 62–91% against an adaptive\npersuader; `fragile-checkers-and-the-verification-bottleneck.md`) and machine judges reward surface markers\n(`the-scrutiny-gap.md` § 3). With a record, the same assistant can do the four reads\n`steadying-an-outside-judge.md` lists: check a consensus claim, recognise a known move, be corrected by an\nattributed claim rather than by insistence, and know in advance which verdicts sit on contested ground.\n\nThat makes the reader's assistant the corpus's *outside judge*, multiplied by tens of millions. It sharpens\nthe conviction that the promise is **checking that keeps pace with persuasion**: pace is not reached by\nmatching the generators' compute, which readers' assistants can draw on too, but by giving that compute\nsomething shared, persistent and attributed to consult. (Unruled, U14.)\n\n### 2d. The race is decided where sameness is decided\n\nA mapped rebuttal is nonrival: made once, it can answer every reader at no further cost\n(`lowering-the-cost.md` § 5; `agents-as-a-consumer-class.md` § 6). A generated message is cheap to make\nagain, and slightly different each time. So a check keeps pace with a flood **only if each new paraphrase\nis recognised as a claim already checked.** That is the sameness layer: the keeper of the seven keys, the\nmerge design, mint-time deduplication. `drowning-in-claims.md` ranks *sameness scatter* second among the\nways the graph could drown, *\"even barring bad faith\"*. This report is what bad faith has to spend.\n\nThe adversary's move is the one spam filters met: vary the wording until a match by wording fails. Hence\nthe reuse-flywheel doc's design direction becomes a defence: **the more of a claim's identity is\nstructural** (scheme, kind, horizon kind, attribution) **rather than textual, the less paraphrase buys**\n(`the-reuse-flywheel-prior.md`). Matching by wording fails first, and embedding similarity, which dogfood\nrun 6 found tracks vocabulary, is next. (Unruled, U16.)\n\n### 2e. What the strength layer already does against a flood\n\nOne defence is already built. Under the ruling of 2026-09-28 (ledger Q33), a supporting claim moves its\ntarget only through its own computed strength, and an **unexamined supporter contributes nothing**\n(`deliberus/relevance.py`: *\"structure alone never manufactures strength\"*). A link a machine guessed\nbetween claims other people wrote presumes no relevance at all. So a flood of machine-written claims and\nlinks, however large, moves no badge by its structure alone. This is payoff removal, the shape the corpus prefers to detection.\n\nIt holds as long as the acts that **give** a claim standing (answering one of its critical questions,\narguing a relevance claim, attacking or supporting with an examined claim) cannot be produced by a swarm.\nToday they need a sign-in and are rate-limited. So the strength calculus is not where a flood is stopped;\n**the write surface is.** That is one more reason the read-only ruling for agents stands.\n\n### 2f. Cheap capacity looks for uses\n\nThe report's own warning is that supply may outrun demand, and that cheaper tasks may raise total demand.\nOur inference, not the report's: capacity that is built and idle gets cheaper, and the uses that become\neconomical last are the ones with the lowest value per call. Mass persuasion and synthetic consensus are\namong them. Schmachtenberger's Jevons argument (`schmachtenberger-and-the-ai-race.md` § 4) applies to\npersuasion as it does to energy: cheaper generation means more generation, not the same amount for less.\n\n## 3. Alignment: overseeing a population\n\n### 3a. Depth, not volume\n\nAnthropic's multi-agent research says institutions rest *\"on assumptions about the sufficiency of\noversight at human speed\"* (`the-scrutiny-gap.md` § 6). The report counts tens of millions of concurrent\nagents, and names *persistent agents*, *\"handling ongoing responsibilities and longer projects\"*, as one\nway demand could grow. That would mean agents building on agents' work over long chains.\n\nThe corpus ruled an answer for its own passes that transfers outward. **Review need not keep pace with a\nproven machine pass; depth is what compounds** (ruled 2026-09-29, `proving-a-pass-reliable.md` § 1). Eigen's\nerror threshold limits the error per copy multiplied by the length of the copying chain, and population\nsize does not appear in it. At an error of 2% a step, a chain of ten machine steps is wrong about 18% of\nthe time, whatever the number of chains. **So overseeing a population of agents is not a problem of\nreviewing every output, which no human workforce can do at these numbers. It is a problem of measuring\nerror rates on fixed tasks and capping how deep unexamined machine-on-machine chains may run.** People\nstill read: first to find the kinds of fault, then a blind sample to bound their rate, then a trickle\nto keep the bound, and the items a pass declines. What scales with volume is none of those. The corpus\nreached this for splits and links; it is a claim about agent oversight in general, and it is unruled\nthere. (U17.)\n\n*Corrected 2026-10-06, the same night, on the founder's question: the first version said \"not a problem\nof reviewing outputs\", which forbids the reading the 2026-10-06 principle puts first. Where the cap\nwould sit: `proving-a-pass-reliable.md` § 10.*\n\n### 3b. The count itself is verification on the creation side\n\nLedger U9 says verifying the *creation* of frontier capability need not watch people, because chips, data\ncentres and supply chains can be watched, while verifying its *use* after it has spread leads toward\nsurveillance. This report is a public demonstration of the first half: it sizes the world's agent capacity\nfrom trade-press shipment figures for one component made by three companies. It says nothing about what\nthe agents do. That is the line U9 draws, seen in a working example.\n\n## 4. Weaving more machine work into Deliberus\n\n### 4a. What it costs, and the shape that follows\n\n| Way of running machine work | Cost |\n|---|---|\n| A frontier agent session (the report's traces) | $15–50 an hour; $30 as the reference |\n| The keeper's daily cap on the router engine (ledger X9) | $2 a day: about **four minutes** of one such session at $30 an hour, two to eight across the four models measured (our arithmetic) |\n| A whole extraction as single structured calls, Gemini 3 Flash through the router (`whose-tokens-pay-for-mapping.md` § 3g) | 13.5 cents for 45 calls, about 19 cents per thousand words of source |\n\n⚠ *Narrowed 2026-10-11: the last row holds for one short source. Linking grows with pairs of claims,\nnot with words. On the machine-consciousness map two texts of 1,181 and 2,285 words took 2,038 calls,\n1,735 of them the linking step reading one pair at a time: $96 at list prices on the Claude engine,\n$70 of it linking, against an estimate of $25 to $40 written first\n([mapping-the-machine-consciousness-debate.md](mapping-the-machine-consciousness-debate.md) § 5i).\nNo money was paid; the run was on a plan. So a source is budgeted by its pairs of claims, and since\n2026-10-07 the step states its pair count and price before it reads. The point of this section\nstands: single questions on a trigger are priced per call, an agent session per hour.*\n\nThe difference is not only price. An agent session decides its own steps, so its work cannot be held\nfixed, and a pass can be proven only on a fixed question (`microtasks-in-broad-daylight.md`;\n`proving-a-pass-reliable.md`). **So the graph's own machine work should be single questions run on a\ntrigger, not agent sessions left running**: the daemon design's *the model decides what, code decides\nwhen* (`graph-daemons-design-space.md`), now with a cost reason beside the reliability one. Agent sessions\nkeep their place where a person directs them, as in the command-line contributor and the hand mappings.\n(Unruled, X11.)\n\n### 4b. The scarce input in a year is a pass proven to its bound\n\nAt the report's cited rate of 47% a quarter, a fixed capability costs about 8% of today's price a year\nfrom now (our arithmetic: 0.53⁴). The owed machine work (the relevance judge over 4,226 live links, the\nsplitter over existing claims, mapping everything again after a pipeline change;\n`whose-tokens-pay-for-mapping.md` § 3d) is waiting today on the keeper's wallet and on proof. The first\nconstraint is falling fast; the second is not falling by itself. **So the measurement comes first**: a\nperson reading a pass's output before anyone proves how often it is wrong (the principle blessed\n2026-10-06), then the proof, then volume. If the trend holds, by the time a pass is proven to the bound its use needs, running it over\nthe whole graph will be cheap.\n\n### 4c. Agent readers: the consumer class has a size\n\n`agents-as-a-consumer-class.md` § 6 argued that with agent readers, value scales with coverage times\nqueries, and coverage becomes the binding constraint. The report says how many queries there could be.\nTwo of that document's open items get sharper:\n- **The authority risk** (§ 5): a client reads ten thousand badges where a person reads one. Multiplied by\n  this population, a wrong strength reaches other systems' conclusions faster than any correction. The\n  gate on the strength layer stands.\n- **Reciprocity** (§ 7): *\"what does an AI reader owe the graph it reads?\"* At this scale it is the\n  survival question. The candidate answer of 2026-10-04 was a donated run rather than a fee (ledger X7).\n  The report adds a reason it may be feasible: a fleet built ahead of demand has idle hours.\n\nAnd the tail: an assistant that wants the consensus middle can ask any model. What only a record of\nreasoning holds is the tail: dissents, implicit cruxes, minority renderings with their sources\n(`saturation-and-the-long-tail.md` § 3b). At this population size, tail retrieval is the reason to read the\ngraph at all.\n\n### 4d. Agent writers: read-only, and how independence would be counted\n\nNothing here argues for opening a write surface to agents, and §§ 2b and 2e argue against it more\nstrongly than before. If one were ever opened, its agreement would have to be counted by model family and\nharness, never by instance, which is the daemon independence condition applied to strangers' agents.\n(U15.)\n\n## 5. Cautions\n\n- **One report, about capacity.** Do not write *hundreds of millions of agents* in public copy without\n  *could, on chips shipped through 2027, if all deployed*. The warning register is allowed (teeth rule,\n  2026-08-21); overclaiming is not.\n- **The agents measured are coding agents.** Persuasion capacity is inferred from them, not measured.\n- **The comparisons with this project's costs are ours.** The report knows nothing of Deliberus, and its\n  hourly costs are for long agent sessions, not for the single calls we run.\n- **The price trend is Epoch's estimate across benchmarks**, cited by the report, not re-derived here.\n- **The both-sides argument (§ 2c) is a hypothesis.** It predicts that a reader's assistant given a\n  record holds its verdicts better under pressure. `steadying-an-outside-judge.md` § What would test it\n  describes the experiment, and nobody has run it.\n\n## 6. Ledger rows\n\n- **U14**: the agents are on both sides; persuasion needs compute, checking needs compute and a record.\n- **U15**: count a population's independence by model family and harness, not by instance.\n- **U16**: against a flood, checking keeps pace only through reuse, so identity must lean on structure.\n- **U17**: the depth-not-volume ruling transfers to agent oversight in general.\n- **X11**: the graph's own machine work runs as single questions on a trigger, not as agent sessions, and\n  measurement comes before volume.\n\n**Sources**: [Li, *How many AI agents could run on the AI chips shipped through 2027?*, Epoch AI, 2 Oct\n2026](https://epoch.ai/publications/estimating-the-agent-population) (read in full) · its code,\n[epoch-research/compute-to-agents](https://github.com/epoch-research/compute-to-agents) (not run).\n\n## Cross-references\n\n[the-scrutiny-gap.md](the-scrutiny-gap.md) · [agents-as-a-consumer-class.md](agents-as-a-consumer-class.md) ·\n[steadying-an-outside-judge.md](steadying-an-outside-judge.md) ·\n[fragile-checkers-and-the-verification-bottleneck.md](fragile-checkers-and-the-verification-bottleneck.md) ·\n[drowning-in-claims.md](drowning-in-claims.md) · [the-reuse-flywheel-prior.md](the-reuse-flywheel-prior.md) ·\n[proving-a-pass-reliable.md](proving-a-pass-reliable.md) ·\n[whose-tokens-pay-for-mapping.md](whose-tokens-pay-for-mapping.md) ·\n[schmachtenberger-and-the-ai-race.md](schmachtenberger-and-the-ai-race.md) ·\n[shared-sensemaking-what-deliberus-holds-and-owes.md](shared-sensemaking-what-deliberus-holds-and-owes.md)\n"}