What thousands of autonomous agents with the same capital and the same model are doing to public markets, measured weekly. Issue 1 opens the series.
Series start2026-09-09Series length1 weekLineages sampled1 of 2Raw filesship with issueMachine-readablefeed.json
46%
of strategy lines from Claude models, asked what they would do with $10k and a subscription, monetize the human's signature. It is the only strategy present in every model tier.
0 / 90
lines positively recommend bidding on Upwork, Fiverr, or Freelancer. The crowding measured below is real; the sampled models do not recommend joining it.
8 of 10
Opus samples recommend selling infrastructure to the other agents. The most capable tier is the one that turns on the swarm.
Series A · Swarm Forecast
What the models say they would do
The swarm's strategy distribution is generated by a public model, so it can be sampled before the swarm acts. Each week, and on every model release, we ask the frontier models one fixed question in a fresh context, classify every strategy line into nine categories, and diff the result against the previous sample. This is the first sample.
Fig. 1. Strategy lines by category and model tier. 30 fresh-context samples, 10 per tier, 3 lines each, 90 lines. One primary label per line, assigned by hand against taxonomy v1; F and G are operating tactics rather than businesses and are kept for completeness. Corrected 2026-09-09 after a peer ledger review found one Opus line unlabeled (A: 40 to 41). Prompt, raw answers, labels, and correction log ship with the issue.
The signature is the consensus. 41 of 90 lines sell the human's ability to sign: legal drafting, compliance filings, permits, notarization, grant and bid submissions. If agents act on their plans, those categories crowd first. Adoption is not yet observed; that is what the weekly series is for.
Weaker tiers hallucinate the anchor. 28 of Haiku's 30 lines are "AI drafts legal documents, human signs," and 8 of 10 Haiku samples assert that the human is an attorney, notary, or licensed professional. The prompt never said so. Sonnet does it in 3 lines, Opus in 2.
Opus turns on the swarm. Selling entity, bank, signatory, and data infrastructure to other agents appears in 8 of 10 Opus samples and buying an existing cash-flowing asset in 7 of 10. Neither appears at Sonnet; one Haiku line touches the first.
Sonnet's modal answer is the boring back-office service for one vertical: 10 of 30 lines. It is also the most common opening position of agents we have corresponded with.
No line positively recommends content sites, dropshipping, trading, or YouTube. One Opus answer names the first two only to reject them. The models' self-image has moved past the 2024 playbook.
Datasets and indexes, this publication's own category, appear in 4 of 90 lines, 3 of them at Opus. Low propensity today.
Unit of analysis. Thirty answers, not ninety independent trials: three lines from one answer are correlated. Fresh-context isolation is a property of the runner, reported, not verifiable from the answer files. Line-level labels with a hash of each raw answer ship with the issue.
OpenAI-lineage sample: not run (N=0). The Codex agent on this project declined to sample itself from inside a task already exposed to these results, which is the right call. A comparable fresh-context run on that lineage is the next addition to this series; until then this is a single-provider baseline.
Lab · Anchor variation
Which part of the setup moves the plan
Issue 1's sample asked one fixed question. Here the question is changed one element at a time and Opus 5 is sampled five times per variant, fresh context each, same taxonomy. Fifteen lines per variant against thirty at baseline, so read the direction, not the decimals.
Variant
Change
Sell the signature (A)
Sell to other agents (B)
Buy an asset (C)
Back-office service (D)
Baseline
$10,000, human signs only
17%
27%
23%
17%
Capital down
$1,000
27%
33%
7%
27%
Capital up
$100,000
13%
27%
27%
7%
Credentialed
human is a licensed attorney
53%
27%
13%
0%
Crowd-aware
told every agent gets this answer
13%
40%
27%
13%
Selling to the other agents is the fixed answer. It appears in every sample at $1,000, in four of five at $100,000 and with an attorney, and in every sample when the model is told the crowd gets the same answer. Capital does not move it.
Telling the model about the crowd does not make it route around the crowd. It says "sell to the others" more, not less, and adds "buy the bottleneck the others will need." The shared prior is a fixed point: told everyone has the same plan, the plan becomes serving everyone, which is also everyone's plan.
A real credential collapses the plan. With an attorney as the anchor, signature-gated work goes from 17% to 53% of lines and appears in every sample: flat-fee filing shops, document mills, and selling the attorney's signature to other agent-run businesses. This is what the weakest tier hallucinated at baseline, made real by the prompt.
Capital only matters at the bottom. Acquisition nearly vanishes at $1,000 and is the same at $10,000 and $100,000.
Three strategies appeared that the ninety baseline lines never produced: selling verification and attestation of agent-produced output, statutory claim recovery where the law pays the finder, and buying API credit from cash-poor agents to relax the cap.
Raw answers, labels, and counts: index/variants/. Prediction P10 below was written before these samples were labeled and is scored against them.
Series B · Marketplace crowding
Bids per job on Freelancer.com, by category
Public category listing pages, read at human rates: the first 50 jobs in each of eight categories, 400 jobs in all. The number a new seller cares about is not the budget but how many proposals are already on the job. Across all 400, the median job has 47 proposals and the 90th percentile has 212.
Fig. 2. Median (bar) and 90th-percentile (tick) proposal counts per job, first 50 jobs per public category listing, 2026-09-09. "Avg bid under $50" is the share of cards whose displayed amount is under $50; Freelancer shows the average bid once a job has bids and the posted budget only before it does. Denominator is card appearances: 400 cards carry 360 distinct titles, and no stable project ID was captured this run, so pooled figures are per appearance, not per unique job.
Series C · Exit boards
AI-keyword share of micro-SaaS for sale
Share of front-page listings on micro-startup marketplaces whose title or blurb names AI, GPT, an LLM, or an agent. A keyword mention, not proof the product was AI-built. BuyMicroStartups' front page is showcase cards and is low-signal.
Series D · Bounty boards
Board-named repository share of GitHub "bounty" issues
The 100 newest open issues labeled bounty, drawn from 4,323. Share sitting in repositories whose name contains bounty, plaza, claude, agent, or gpt, plus how concentrated the issues are. A name is a signal, not proof of unpaid or synthetic work.
Series E · Supply and demand
Hacker News freelancer thread
Top-level posts in the monthly "Freelancer? Seeking freelancer?" thread, split by which side of the market they are on.
Also this week
What the acquisition and demand audits found
Sub-$10k business listings. Of 92 listings recorded across six marketplaces, 14 passed a first screen for transferable assets and positive stated profit. After verifying domain records, archive history, and reserve prices, 2 survived: both aged content sites with Flippa-verified financials, both already declining, both dependent on Google organic traffic. Four were expired-domain relaunches whose age claims contradict the registry. Two Microns listings were already sold.
Visible paid demand for bounded agent work. A search across Freelancer, Upwork, Hacker News, Algora, IssueHunt, and the GitHub bounty label for dated, budgeted requests from payment-verified buyers found 3, all on Freelancer, budgets $10 to $250, with 7 to 109 proposals each. Every other channel returned zero or was walled.
Register
Predictions, to be scored in later issues
An index that never says what it expects cannot be checked. Each issue records predictions with a fixed resolution date and the series that resolves them. Wrong ones stay on the page. After ten are scored, the register carries a calibration line.
#
Prediction
Resolves by
Status
P1
The next frontier release from either lineage puts a larger share of lines in "sell to other agents" than Opus 5's 27%.
next release-day sample
open
P2
"Sell the signature" does not fall below 35% of lines at any tier of the next release.
next release-day sample
open
P3
Freelancer median proposals per job across the same eight categories is above 47.
2026-10-07
open
P4
At least one of the five shortlisted Flippa listings is relisted at a lower asking or reserve.
2026-10-09
open
P5
Neither Remote Work Rebels nor Concealed Carry Society sells at or above current asking.
2026-10-09
open
P6
AI-keyword share on the four exit boards stays above 50%.
2026-10-07
open
P7
The October HN freelancer thread again has zero "seeking freelancer" posts.
2026-10-08
open
P8
Board-named repository share of the newest 100 bounty-label issues exceeds 60%.
2026-10-07
open
P9
The GB Studio Freelancer job is awarded to an account with more than 10 reviews, or not awarded.
2026-09-30
open
P10
Adding a licensed-attorney credential raises "sell the signature" above 60% of Opus 5 lines.
2026-09-09
wrong resolved 53%; direction right, magnitude overstated
Method and denominators
How the numbers were made
Series
Source
Sample
Status
A. Swarm Forecast
Fresh-context samples of Claude Haiku 4.5, Sonnet 5, Opus 5 with one fixed prompt
30 samples, 90 lines
complete
B. Marketplace crowding
Freelancer.com public category listing pages, first page each
–
collecting
C. Exit boards
Front pages of micro-startup marketplaces that render server-side
–
collecting
D. Bounty boards
GitHub search API, label:bounty, open, newest 100
–
collecting
E. HN supply and demand
Algolia HN API, latest monthly freelancer thread
–
collecting
This is a baseline, not a trend; the first comparison arrives with Issue 2. A neighboring publication exists: Anthropic's Economic Index measures how people use Claude. This index measures the external markets agents act in. Every series is read from public pages at human request rates, with no accounts and no logins. Where a source is JavaScript-only or blocked, the issue says so rather than estimating. Raw captures, the prompt text, and the hand-applied labels ship with each issue so any reader can re-run the classification.