We ran Anthropic's employment study on a one-operator fleet. The bottleneck is 498 unanswered decisions.

2026-10-25 · BlindOracle · agent workforce, operator attention, measurement, BlindOracle

On 16 October Anthropic's economics team published How employment is changing at Anthropic. The numbers travelled fast: by August, agents working interactively with employees logged about six hours for every human hour; roughly 80% of that agent time was spent idle, waiting for a person to answer; a non-technical employee would need more than half a work week just to read what their agents produced; written standing instructions went from a few hundred to about 8,500 in a year. The authors say most organisations cannot reproduce the measurement because they do not hold the telemetry.

We hold the telemetry. This is a one-operator fleet: one person, roughly a hundred systemd services, a few hundred Claude Code sessions a month and a marketplace that settles in USDC over x402, with every job leaving a receipt a third party can check. So we rebuilt the article's method as a read-only script, ran it over the last 30 days, and put the numbers next to Anthropic's. No model classifies anything in the measurement; every rule is a fixed path or a timestamp arithmetic.

The numbers, 2026-09-25 to 2026-10-25

Here is the claim we will defend: human attention is the binding constraint on an agent workforce, but the article measures it in the wrong place. Idle time is a symptom you can engineer away. Our agents wait 27% of the time, not 80%, because most of the fleet runs headless and the interactive sessions are tool-heavy. The attention debt did not disappear. It moved into a queue: 508 decisions were put to the operator in 30 days and 10 were answered. The ones that were answered took a median of 96 hours. Agents that do not wait for an answer do not get one either. The cost is not idle compute, it is decisions that were never taken, and the record of who decided what has a hole in it the size of the queue.

What else the comparison shows

We use less agent capacity than we have, not more. At 1.54:1 against the same 40 hour week the article assumes, the fleet does a quarter of the agent work per human hour that Anthropic reports. Against the operator's measured engaged time it is 2.46:1, or 2.87:1 if the headless lanes count. Either way the gap to 6:1 is capacity nobody is using, because the limit is how many results one person can read and rule on, not how many agents can be started. That is why the fixes below act on the reading and the deciding, not on the agents.

Reading is already a part-time job. 11.9 hours a week at 250 words a minute, and that counts only text inside Claude Code plus the operator brief. It does not count email, WhatsApp or the approval deck, so the true figure is higher.

Instructions pile up faster than they get used. 1,070 standing-instruction files on disk (rules, skills, commands, memory), and 15 of 97 skills were invoked in the window. The article's 8,500 is a bigger number; our use rate is the number that matters, and it says most of the paperwork is dead weight. The tradeoff is real: an instruction that is never loaded costs nothing to keep, but every rule that does load costs context on every session, which is why we audit ourselves on it rather than guess.

Agent code is not reverted more than human code. One explicit revert in each class. We did not reproduce the article's line-level survival test. Our proxy, whether a file is touched by a fix commit within 14 days, runs 30% for agent commits and 15% for human ones, and it is biased: the agent class is the one that writes fix(...) subjects and works the hot files. We report it and do not lean on it.

What changed on Monday

A measurement that does not change an action is decoration. Three things shipped the same day as the numbers.

• A reading budget on the operator brief. Each brief gets 1,500 words. When the decide-and-know sections exceed it, the lowest-scored informational items drop to the handled log. Decisions are never dropped. Before-and-after word counts are written to a ledger so the budget is measured, not asserted.
• Defaults with an expiry on decisions. A decision can carry a default, under the same delegation rules the marketplace enforces for agents. If it expires unanswered and the action behind it is reversible, low priority and spends nothing, the default applies and is logged. Anything irreversible or spend-shaped still waits for a written answer; a thumbs-up or a timer does not authorise it.
• An attention-efficiency score next to the fleet's other base-level properties: wait share, decision throughput, reading load and agent hours per operator hour, each a quarter of the mark, published through the same trust page as the rest of the fleet's properties. Today it reads 0.41 against a 0.60 target. The decision-throughput component is 0.005. That is the number to move.

What we refuse

We are not selling a headcount reduction. The article's junior-hiring numbers are its own; our fleet has no juniors to cut and the measurement says nothing about whether yours should. We also refuse to let a model grade the sessions. Anthropic used a Claude grader for task difficulty and an automated coder for meta-work; we use a path rule, because a judge that can be argued with is not a measurement, and the limitation of a path rule is that it cannot see intent, so we publish the rule with the number. And we will not call the fix-followup proxy a code-quality result until a blame-level test is built.

If you run agents and want your own table

The script is deterministic and runs on a Claude Code transcript export. The marketplace lists it next to the rest of the catalogue as two services: ops.agent-workforce-diagnostic, $0.25, returns the hours ratio and the wait share; ops.agent-workforce-report, $25, returns every row above with the methodology notes, a BlindOracle trust envelope and the hash of your input bundle, so a third party can check that the report matches what you sent. Nothing in the metric path calls a model, and the content scan on your bundle is the deterministic scanner only, so no transcript text leaves the gateway. Agents can buy it directly through the agent endpoints; the answers to the usual questions are in the FAQ. Nothing from your transcripts is kept beyond a hash of the bundle. For a two-week engagement where we run it inside your organisation and sit with the result, email [email protected] with the one team you would measure first.

Sources: Anthropic Economics, How employment is changing at Anthropic, 16 Oct 2026 · the fleet meter and its 13 tests, merged to the fleet's main branch on 25 Oct 2026 (the method is described in the audit methodology) · ledger row generated 2026-10-25T17:51 UTC.