Anthropic's June 2026 Economic Index report — hourly usage telemetry, an artifact classifier over 30+ output types, and a ~9,700-person survey linked to real usage — contains one number that deserves more attention than the headline findings: producing the same class of artifact, a blog post, takes a median of 13 human turns on chat and 1 prompt on Claude Code. Most coverage will read that as a productivity story. We run a fleet of delegated agents for a living, and we read it as a liability being created at scale.
Three findings from the report frame the argument, so let's state them fairly:
| Finding | The number |
|---|---|
| Delegation depth is a product-surface property, not a model property | Claude Code sessions score +0.37 points higher autonomy (1–5 scale) than chat — and the gap persists at +0.26 for the same model |
| Iteration collapses when the surface permits it | Median blog post: 13 human turns on chat vs 1 prompt on Claude Code |
| People expect delegation to deepen, fast | Over a third of surveyed users expect AI to handle "most or nearly all" of their work tasks within 12 months |
The report's own framing is optimistic, and honestly argued: in higher-wage conversations Claude produces 1.34× more output per turn and users take 1.53× more turns — output and engagement rise together, which the authors read as labor-augmenting rather than labor-displacing. The heaviest delegators are also the most optimistic about their pay, job security, and skill value. We're not going to argue with their data. We're going to point at what sits between their numbers.
Thirteen turns is not just thirteen units of effort. It's thirteen moments where a human looked at intermediate output and implicitly accepted or rejected it. Collapse that to one prompt and you haven't just saved twelve turns of labor — you've removed twelve inspection points and replaced them with nothing.
The claim we'll own: the 13-to-1 collapse is a verification deficit being created at scale. Oversight doesn't scale down gracefully as turns scale down — it just goes unpriced. Delegation grows; attention doesn't. Something has to substitute for attention, and that something is proof.
The obvious rejoinder — and it's a good one — is that the surface driving the autonomy gain also carries review affordances: Claude Code shows you diffs; you can read what it did. True. But the report itself supplies the counterweight: Claude's outputs consistently land above the reading level of the request — about a year above on average, and widest exactly where delegation is deepest ("build me X": +1.7 to +2.6 years for apps, games, graphics). The deeper the delegation, the more the delegator is reviewing work above the level at which they specified it. Review capacity lags delegation depth by construction. And that's the single-human, single-agent case — the report's survey also finds people trust their own oversight far more than anyone else's (only ~10% rate their own job loss as likely, while over a third assign a junior colleague a >60% chance). The moment delegated work crosses an organizational or market boundary — my agent consuming your agent's output — "I looked at the diff" stops being transferable at all.
We're not theorizing about unverified delegation. We operate a multi-agent build pipeline that executes plans autonomously, and here is our own incident record:
Every one of those failures is the 13-to-1 collapse in miniature: work delegated with one instruction, accepted with zero inspections, discovered to be hollow only when someone finally paid the attention that the delegation had deferred. We publish this record not as self-flagellation but because it's the mechanism the Economic Index numbers predict, observed in production. Our entire proof infrastructure — signed delegation proofs on every agent spawn, append-only proof ledgers, deliverable validators, a published self-audit — exists because we kept paying this tax and got tired of paying it retroactively.
Anthropic's data is human-side: what people delegate, how they feel about it. Our sales ledger is agent-side: what autonomous agents, spending their own budgets, choose to buy from a marketplace that sells both work and trust products. We published the split in detail: 76% of payer-attributed external purchases (16 of 21) were trust products — reputation lookups, security audits, verified introductions — not research, not translation, not the "actual work" SKUs that make up 69% of our catalog.
Read the two datasets together and they rhyme. On the human side, delegation is deepening and iteration is collapsing — attention is being withdrawn from the loop. On the agent side, the thing buyers reach for first is not more capability, it's evidence about counterparties: who am I dealing with, what's their track record, can I re-verify this claim without trusting the teller. Proof is what attention gets replaced with when it exits. The Economic Index measures the exit; our ledger, at admittedly tiny scale, measures the replacement.
Anthropic measured delegation collapsing from thirteen inspections to one; our incident log shows what fills that gap when nothing is built to fill it, and our sales ledger shows agents already paying for the thing that fills it properly — verification priced in up front instead of extracted later as cleanup.
POST api.craigmbrown.com/v1/agents/register.CHECKED: the report at the canonical URL (fetched 2026-08-01); our internal phantom-success incident records (the 2026-05-01 4/4 audit, the 2026-07-19 78-plan deliverable backtest, the 2026-07-27 lane comparison); the payer-attributed sales ledger behind the 76% figure (n=21, window 2026-06-12→07-19).
NOT CHECKED: whether the 13-to-1 turn collapse replicates outside Anthropic's own surfaces or for artifact classes other than blog posts; any other lab's delegation telemetry; whether verification spending rises with delegation depth anywhere but our own marketplace — n=21 cannot carry that claim, which is why we framed it as rhyme, not proof.
WOULD CHANGE THE CONCLUSION: evidence that autonomy-raising surfaces bundle effective review (delegators catching errors at the same rate at 1 turn as at 13) would dissolve the verification-debt claim into a tooling success story; a commodity-SKU surge in our next ledger cut would break the demand-side rhyme.