Polsia
The famous number is not ARR
The counter was 12 times the last 30 days of all money received. Subscriptions were 47.5% of it in May 2026 and 60.9% in August. Advertising budgets and the sales of customers' companies pass through Polsia; most of that money is not Polsia's revenue. This page is on Aiolum as a warning about how a public revenue claim can mislead, not as a model to copy. Read the section "What to watch" before the numbers.
Polsia sells AI agents that build and run a small online company for a customer. An AI agent is a program in which a language model decides the next action and performs it with tools. Examples of tools are a code editor, an email account and an advertising account. The customer gives an idea. Polsia creates a website, an inbox and a payment account. One agent does one task every night and reports by email. The customer pays a monthly subscription and buys credits for extra tasks.
Ben Broca, who also uses the name Ben Cera, launched it alone in December 2025. He is a former manager at CloudKitchens. In February 2026 he let the product answer investor emails in public, and he showed a live counter of the company's run rate. A run rate is the revenue of a recent period, multiplied to a full year. Here it is the last 30 days times twelve. The counter passed US$1 million on 28 February 2026 and US$10 million in May 2026. In May he announced an investment of US$30 million at a US$250 million valuation. He has no employees.
The famous number is not annual recurring revenue (ARR), which is the monthly subscription revenue times twelve. The counter also annualized one-time credit purchases, advertising money that customers send through Polsia to Meta, and the sales of the customers' own companies. On 23 May 2026, subscriptions were 47.5% of the counter. This page therefore uses the subscription figure from the company's own public data: US$774,984 a month on 27 August 2026. Every number here is reported by the company, grade C. No payment processor, auditor or journalist has checked the payment records. The founder says that churn is about 50% in the first month. Churn is the share of paying customers who cancel in a period.
The numbers, graded
Each row shows the evidence grade and the source numbers from the list at the end of this page. The grades are explained in How Aiolum verifies.
| Fact | Value | Grade | How we know | Sources |
|---|---|---|---|---|
| Subscription revenue per month, 27 August 2026 | US$774,984 | C | The field mrr on the company's public dashboard, in the copy that the Internet Archive saved on 29 August 2026. The dashboard lists 20,382 paying users on that date. | [1] |
| Subscription run rate, 27 August 2026 | about US$9.3 million a year | calc | Calculated: 774,984 times 12 equals 9,299,808. The dashboard's own field arr shows 9,299,809. This is the headline figure on this page. | [1] |
| Counter shown to the public, 29 August 2026 | US$15,532,004 | C | The field arr_usd, which the live page showed as the run rate. It is 12 times the money received in the last 30 days, from every source. | [1] |
| What the counter contained, 29 August 2026 | subscriptions 60.9%, one-time purchases 27.0%, customers' money 12.0% | calc | Calculated from the 30-day fields: subscriptions US$788,489, credit packs US$332,370, domains US$17,619, advertising budgets US$120,279, sales of customers' companies US$35,576. Their sum times 12 equals the counter to the dollar. The rounded shares add to 99.9%. Under the terms, 80% of an advertising budget goes to Meta, and 97% of a customer's sales belongs to the customer. | [1] |
| What the counter contained, 23 May 2026, one day after the funding news | subscriptions 47.5%, one-time purchases 28.9%, customers' money 23.6% | calc | Calculated from the same fields. Counter US$9,762,253; subscriptions US$386,643 a month; credit packs US$162,805; boosts, a paid extra, US$72,157; advertising budgets US$162,387; sales of customers' companies US$29,529. | [1, 9] |
| Subscription run rate, 22 May 2026 | US$4,630,500 | C | The dashboard's own field arr: US$385,875 a month times 12, from 5,830 paying users. On that day the founder announced an annual run rate close to US$10 million. | [1, 7] |
| Date when the counter passed US$1 million | 28 February 2026 as first shown; 6 March 2026 in the later version | C | Copies saved in March 2026 show US$1,098,344 for 28 February. Copies saved from 15 April 2026 show US$590,378 for the same day and US$1,086,102 for 6 March. The company changed its history between 29 March and 15 April. This page found no explanation. | [1, 2] |
| Subscription run rate when the counter passed US$1 million | US$443,208 on 1 March 2026 | C | An independent developer copied the dashboard data that day: US$36,934 a month from 592 paying users. On 6 March he copied US$923,232 from 1,193 paying users. No archived copy exists for those two days. | [8] |
| Time from launch to US$1 million on the counter | 74 days, not 30 | calc | Calculated: 16 December 2025, the first payment in the history, to 28 February 2026. The claim of 30 days counts from 26 January 2026, when the counter showed US$20,585. That is 33 days. | [1, 2] |
| The investor's version | US$1 million ARR within one month of launch, US$3 million one month later | C | True Ventures, 23 March 2026. The archived counter shows US$1 million 74 days after the first payment and US$3 million 8 days after that. The post does not say how the investor checked the figures. | [5, 1] |
| Paying users | 592 on 1 March, 5,830 on 22 May, 20,382 on 27 August 2026 | C | Dashboard field paying_users. The Wall Street Journal reported 10,000 paying customers on 29 July 2026. The dashboard showed 11,761 on 24 July. | [1, 6, 8] |
| Subscription revenue per paying user | US$66 a month in May, US$38 a month in August 2026 | C | Dashboard field arpu. The first plan cost US$49 a month. A reviewer reported an entry plan of US$20 a month by August 2026. | [1, 3, 10] |
| Customers who leave (churn) | about 50% in the first month, about 50% again in the second | C | Founder, 27 May 2026. The dashboard field for 30-day paid churn showed 48% on 22 May (2,689 of 5,598) and 81.5% on 27 August 2026 (13,712 of 16,820). Polsia's own email agent told one interviewer that 85% of customers stay in month two. The two statements do not agree. | [4, 1, 12] |
| Cost of AI models per day, 18 March 2026 | US$27,272 | C | Dashboard field daily_ai_cost. In May 2026 the founder said that his bill from Anthropic was about US$1 million a month, and US$1.5 million in the last month. | [1, 4] |
| AI cost compared with subscription revenue, 18 March 2026 | about 4 times | calc | Calculated: 27,272.11 times 30 equals 818,163 a month, against subscription revenue of US$205,100 a month. The field showed US$13,624 a day on 22 May and US$3,192 a day on 11 July 2026. In that period the founder moved to cheaper models. | [1, 6, 15] |
| Funding | US$1 million in mid-2025; US$30 million at a US$250 million valuation, announced 22 May 2026 | C | Founder statements. Press reports name Sound Ventures, True Ventures, Offline Ventures, Adjacent, Tekton Ventures, Drysdale Ventures and Vaynerfund. True Ventures confirmed that it invested. This page found no investor statement of the amount. | [7, 4, 16, 5] |
| Valuation compared with the subscription run rate of 22 May 2026 | 54 times | calc | Calculated: 250,000,000 divided by 4,630,500 equals 54.0. Against the counter of US$9,762,253 it is 25.6 times. | [1, 7] |
| Employees | none | C | Founder, and the Wall Street Journal on 29 July 2026. He told Fortune that a law firm and an infrastructure company work for him as outside partners. | [6, 13, 4] |
| Customer companies that are still active, 29 August 2026 | 24,187 of 386,353 created (6.3%) | C | Dashboard fields companies and total_companies; the share is calculated. On 23 May 2026 it was 7,704 of 123,493 (6.2%). | [1] |
| Sales of all customer companies together, 30 days to 29 August 2026 | US$35,576 | C | Dashboard field user_company_30d. This money belongs to the customers; Polsia keeps a fee of 3%. | [1, 18] |
| Average sales per active customer company in those 30 days | US$1.47 | calc | Calculated: 35,576.24 divided by 24,187. | [1] |
| Customer companies that earned at least one dollar | about 10% | C | Founder, 27 May 2026. The best company he had seen made US$3,000 to US$4,000. In February 2026 the best one made about US$50 a month. In April 2026 he told Rest of World that about one-sixth had revenue. | [4, 3, 14] |
| Independent proof of the revenue | none public | C | Not verified by anyone outside the company. On 27 May 2026 an interviewer asked the founder to show his Stripe account. He declined and said that he would post it on X; this page did not find such a post. TrustMRR has no Polsia profile. The one Polsia-hosted company listed there shows US$0.00 of verified revenue. | [4, 19, 17] |
| Public dashboard | online from February to September 2026, gone by 12 October 2026 | C | The Internet Archive holds 24 copies from 22 February to 29 August. A reviewer read it on 15 September and reported subscriptions of US$915,309 a month inside a counter of US$18.71 million. On 12 October the address returned an error. | [1, 10] |
| Trustpilot score | 1.8 of 5 in August, 2.9 on 15 September, 3.3 on 2 October 2026 | C | Reported by two reviewers who sell competing products. On 2 October there were 347 reviews, 30% of them with one star. Trustpilot blocked the download for this page. | [10, 11] |
| Time from launch to the August 2026 figure | 8 months | calc | Calculated: December 2025 to August 2026. | [1, 4] |
How to read this table: a grade C value has one of three origins. It is a founder statement, a figure from the company's own dashboard, or a press or reviewer report that did not check payment records. A "calc" value was computed on this page from the rows above it. No value is grade A or B. The dashboard figures come from copies that the Internet Archive saved on the dates shown.
| Plan | Per month, billed yearly (US$) | Per month, billed monthly (US$) | Limit |
|---|---|---|---|
| Standard | 20 | 1 company, 20 credits a month, no Meta adverts | |
| Pro | 50 | 3 companies, 50 credits a month, Meta adverts | |
| Max | 200 | 10 companies, 200 credits a month | |
| Max 2 | 500 | 25 companies, 500 credits a month | |
| Max 3 | 1,000 | 50 companies, 1,000 credits a month | |
| Extra task credits | 19 to 999 | 15 to 1,000 credits a month, in 7 steps |
Polsia has no public price page. Prices appear after sign-up, and the reviewer found two different price lists inside the product on the same day. The yearly column is empty because the product shows the yearly option only as two months free. One credit pays for one agent task. On top of the plan, Polsia keeps 3% of every payment that the customer's company receives and holds the money for 31 days. It also keeps 20% of every advertising budget. The terms of June 2026 still took 20% of customer payments. The first plan, in February 2026, was US$49 a month.
The timeline
From a first payment of US$19 to US$775,000 a month in subscriptions in 8 months, next to a public counter that always showed more. Each entry carries its grade.
- Mid-2025CBroca raises a pre-seed round of US$1 million. During 2025 he builds four products alone. Three make nothing. [4]
- October 2025CHe starts the product that became Polsia: a company that makes companies. The name is AI slop, spelled backwards. He chose it on purpose. [4, 12]
- 16 December 2025CThe dashboard history starts on this day with a run rate of US$228, which is one subscription of US$19 a month. The founder dates the launch to the middle of December. [2, 4]
- 26 January 2026CThe counter shows a run rate of US$20,585. [1, 3]
- 17 February 2026CHe posts on X that the product will raise its own money. It answers investor emails from his inbox for 14 days, next to a public live dashboard. The counter shows US$208,516 that day. [7, 1]
- 26 February 2026CIn a recorded interview the host reads US$693,608 on the live dashboard. The founder says that the number counts all cash that enters the platform. The best customer company makes about US$50 a month. [3]
- 28 February 2026CThe counter passes US$1 million. The next day the dashboard data show a subscription run rate of US$443,208 from 592 paying users. [1, 8]
- 7 March 2026CHe posts a run rate of US$2.5 million, up from US$1 million in one week. The subscription run rate on 6 March is US$923,232. [7, 8]
- 9 March 2026CAn independent developer publishes two findings: the live page exposes customers' chats and personal data, and the counter is about twice the subscription figure. [8]
- 17 to 18 March 2026CThe founder writes that the past week almost broke him: every supplier reached its rate limits. The dashboard shows AI costs of US$27,272 a day and subscription revenue of US$205,100 a month from 3,218 paying users. [7, 1]
- 23 March 2026CTrue Ventures announces its investment and repeats the US$1 million and US$3 million figures. [5]
- April 2026CBetween 29 March and 15 April the company replaces the history of the counter with lower values for February and March. [1, 2]
- 29 April 2026CRest of World reports on a factory trainer in China who pays US$199 a month, a quarter of his salary. His Polsia company has 7 sign-ups and no payment. [14]
- 22 May 2026CHe announces US$30 million at a US$250 million valuation and a run rate that approaches US$10 million. The subscription run rate that day is US$4,630,500. [7, 1]
- 27 May 2026CIn a second interview he confirms the run-rate method and a churn of about 50% in the first month. He gives a model bill of about US$1 million a month. He declines to show his Stripe account. [4]
- 29 July 2026CThe Wall Street Journal reports 10,000 paying customers and a path to US$10 million of revenue in 2026, with no employees. Broca says that he lost money on many accounts and moved to open-source models. [6]
- 27 August 2026CThe dashboard shows subscription revenue of US$774,984 a month from 20,382 paying users. The public counter shows US$15.5 million two days later. [1]
- 15 September 2026CA reviewer reads US$915,309 a month in subscriptions and 29,132 active companies of 484,592 created. The terms now take 3% of customer payments, not 20%. [10, 18]
- 12 October 2026CThe address of the public dashboard returns an error, and the app no longer has a live page. [1, 18]
How it grew
Four causes, in the order they mattered. Each one is a reusable lesson.
Distribution
The fundraising was the advertisement. On 17 February 2026 the founder posted that the product would raise its own money. He opened a live page with the revenue counter and the agents' work. The counter went from US$208,516 that day to US$1,098,344 eleven days later. He says that each post raised the numbers, and that the higher numbers brought more investors and more posts.
No step between idea and company
A new user types an idea or presses a button that chooses one. Before any payment, Polsia makes a company name, a market report, a live website and an inbox, and posts about the company on X. The founder says that each extra question would lower the share of users who pay. On 22 May 2026 the dashboard showed 3,767 new sign-ups in one day and 21.7% of trials that became paid, 4,467 of 20,548.
A counter that includes customers' money
The live number was 12 times the last 30 days of all money received. That money included subscriptions, one-time credit packs, advertising budgets that go to Meta, and the sales of customers' companies. When the interviewer asked, the founder described this method himself. From March to May 2026 the number that people shared was about twice the subscription revenue. The counter was correct as a sum, but investors and press called it ARR, and it was not.
Credits on top of the subscription
The subscription pays for one automatic task each night. A customer who wants more work at once buys credits, at about US$1 to US$2 per task. On 29 August 2026 one-time credit packs were 25.7% of all money received. The cost is on the same side. The founder says that a task cost him about US$2 with Claude. A task cost US$20 to US$30 when an agent repaired a large code base.
How it was built
These details come from founder interviews, press reports and the company's terms. They are grade C. [3, 12, 15, 16, 6]
- Built on Anthropic's Claude with the Claude Agent SDK. An SDK (software development kit) is a set of ready code; this one is for programs in which the model calls tools. In February 2026 the agent that plans strategy ran on Claude Opus 4.6.
- Each customer company receives a web server, a database, a GitHub account, an email address, a Stripe payment account and a Meta advertising account. Polsia manages all of them, so the customer connects nothing.
- A chief-executive agent runs every night. It reads the state of the company, chooses one task, gives it to an engineering, marketing or support agent, and emails the customer a report.
- The advertising agent writes the advert and generates a video of an invented person with OpenAI's Sora 2. It uploads the video to Meta and stops adverts that perform badly. The customer sets a budget from US$10 a day.
- Suppliers that the founder named in May 2026: Sapiom, Blaxel, Anchor Browser, AgentMail, Stripe, Render, Anthropic, OpenAI, Meta, X, GitHub, Postmark and Amazon Web Services.
- The founder writes Polsia itself with Claude Opus and uses OpenAI Codex as a second reviewer for bug fixes. His own tools cost about US$800 a month. He says that he no longer reads the code.
- By June 2026 the founder had moved tasks from Anthropic's models to open-source models on rented graphics processors. He says that the price was about 100 times lower.
Tools named in the sources
Each one opens its page in the Tools tab, with every story that uses it.
The dead that came before
Five companies in the Loot Drop database promised to do the work of a small company for a customer who could not do it alone. The values below are from that database and are not independently verified.
Builder.ai
The promiseCustom apps for founders who cannot code, built by AI from reusable parts.
Why it diedEach US$20,000 project cost US$25,000 to US$35,000 to deliver, and offshore developers did most of the work that was sold as AI.
Metigy
The promiseAI marketing advice for small businesses that cannot pay a marketing team.
Why it diedIt gave advice on a dashboard. Small businesses wanted the work done for them, and they left within 60 to 90 days.
Dukaan
The promiseAn online store in 30 seconds for small merchants.
Why it diedIt signed more than 5 million merchants. Most sold very little, could not pay even US$10 a month, and left.
Zirtual
The promiseVirtual assistants on a monthly subscription for busy founders.
Why it diedIt hired assistants faster than revenue grew, could not pay salaries in August 2015 and closed without notice.
Trada
The promiseGoogle advertising for small businesses, managed by outside experts.
Why it diedHuman experts did not scale at a profit, and Google automated the same work inside its own tools.
What changed
Each of the five sold the work of a company function to people who could not do that work themselves. Each found that the work cost more than those customers could pay. Builder.ai used human developers behind an AI label. Zirtual and Trada used human staff. Dukaan and Metigy gave tools to customers who earned too little to stay.
The founder says that the models of late 2025 were the first that could use any tool and stay on a task without a person. Polsia does with a model what Builder.ai did with offshore staff, so one person can serve 20,000 paying users. The old problems did not go away. In March 2026 the model bill was about four times the subscription revenue. The founder says that half of new customers leave in the first month. Like Dukaan's merchants, most Polsia companies earn almost nothing: US$1.47 each in 30 days, on average. The founder says that the US$30 million lets him accept low margins and offer a free version.
Can one person copy it now?
The public fundraising, no. The measurement lesson, yes.
The growth came from one founder's public fundraising on X and a pre-seed round of US$1 million. The model bill was about US$1 million a month before revenue covered the costs. Few people can repeat that. What a reader can copy is smaller. Remove every step between an idea and a first result, and charge for usage on top of a subscription. A reader can also do the opposite of the counter: report subscriptions, one-time sales and customers' money as three separate numbers.
Where it is still open: Agents that do one function well for a small business that already has customers, for example bookkeeping, appointments or adverts for one trade. Such a customer has revenue to protect and a reason to stay.
- 1Let a new customer see a finished result before the first payment, with no setup.
- 2Sell a low subscription for the automatic work and credits for work on demand.
- 3Measure the cost of each task from the first day. Polsia's cost per task rose when customers' code bases grew.
- 4Publish recurring revenue, one-time revenue and pass-through money separately, with churn next to them.
What to watch
The risks are the mirror image of the causes above.
The famous number is not ARR
The counter was 12 times the last 30 days of all money received. Subscriptions were 47.5% of it in May 2026 and 60.9% in August. Advertising budgets and the sales of customers' companies pass through Polsia; most of that money is not Polsia's revenue.
Company-reported only
No payment processor, auditor or journalist has confirmed the revenue. The founder declined to show his Stripe account in an interview. The company changed the history of its counter once, to lower values, and removed the public dashboard by October 2026.
Half of new customers leave in the first month
The founder gives about 50% churn in month one and about 50% again in month two. The dashboard field showed 81.5% for the 30 days to 27 August 2026. A run rate assumes that the current month repeats twelve times, so new sign-ups must replace the customers who leave.
Costs were above revenue
In March 2026 the model cost was about four times the subscription revenue. The founder told the Wall Street Journal that he lost money on many accounts. The business now runs on cheaper open-source models and has US$30 million of investment. This page found no public figure for profit or loss.
Most customer companies earn nothing
The founder says that about 10% of companies earned at least one dollar. All customer companies together sold US$35,576 in the 30 days to 29 August 2026. A customer who pays every month and earns nothing has a reason to cancel.
Complaints and privacy
Reviewers report a Trustpilot score of 3.3 of 5, with 30% one-star reviews. The complaints are tasks marked as done but not delivered, credits used on failed tasks, and refunds refused. In March 2026 the live page exposed customers' chats and personal data. Agents have sent emails that customers did not approve.
Sources
The grade next to each source is the best grade that the source can support. "Read in full" means that the text of the page was downloaded and the numbers were checked in it.
- C Internet Archive: 24 saved copies of Polsia's public dashboard data, 22 February to 29 August 2026Read the saved data of 22 of the 24 copies in full, with a script. The company published these figures itself. The archive proves what the company showed on each date, not that the payments exist. The live address returned an error on 12 October 2026.
- C Internet Archive: saved copy of the dashboard's daily history, 12 August 2026Read in full. Daily values from 16 December 2025 to 11 August 2026, in the later and lower version of the series.
- C Mixergy, 2 March 2026, interview by Andrew Warner: This AI generates $689KRead the transcript in full. Recorded on 26 February 2026; the host read the live dashboard during the call. Source for what the counter includes, the first price, the 20% share and the nightly agent.
- C Mixergy, 27 May 2026, interview by Andrew Warner: Polsia: AI Agent + Zero Employees = $10M Run RateRead the transcript in full. Source for the run-rate method in the founder's words, the churn, the share of companies with revenue and the Anthropic bill. Also the source for the pre-seed round, the launch month and the Stripe request.
- C True Ventures, 23 March 2026, Tony Conrad: Polsia: The One-Person Company Is No Longer a MetaphorRead in full. An investor's announcement. It repeats the US$1 million and US$3 million figures and does not say how they were checked.
- C The Wall Street Journal, 29 July 2026: The Rise of Million-Dollar Companies With Just One EmployeeThe page needs a subscription. Read the passages on Polsia in the copy that archive.ph saved on 28 September 2026. The article does not say that the reporter saw payment records.
- C Ben Cera on X, 22 May 2026: Polsia just raised $30M at a $250M valuationRead through api.fxtwitter.com, together with his posts of 17 February, 7 March and 17 March 2026. Source for the funding announcement and the run-rate claims on those dates.
- C SiteBloom, 9 March 2026: Polsia claims a $3M ARR and streams all user chats via public URLRead in full. An independent developer copied the dashboard data on 1 and 6 March 2026. Source for the subscription figures on those dates and for the exposed customer chats.
- C Polsia, the receipts: an anonymous analysis by the X account NotOnKetamine, May 2026Read in full. A hostile analysis of the dashboard data of 22 May 2026. Its arithmetic matches the archived copy. Its claims about human staff were not verified and are not used.
- C Preuve AI, updated 15 September 2026: Polsia Review (2026): Is It Legit? An Honest Founder ReadRead in full. The author sells a competing product and says so. Source for the plan prices, the dashboard values of 15 September and the Trustpilot scores.
- C Crevio, 2 October 2026: Is Polsia Legit? An Honest Review With the Founder's Own NumbersRead in full. Crevio is a competitor. Source for the Trustpilot score on 2 October 2026 and the fee summary.
- C Lean AI Leaderboard, 8 May 2026, Henry Shi: How a Solo Founder Cloned Himself With AI That Now Runs 6000 Companies and Hit $6M+ ARRRead in full. Most answers came from Polsia's own email agent. Source for the founder's background, the Claude Agent SDK, the founder's own tools and the retention claim.
- C Fortune, 26 March 2026, Allie Garfinkle: The one-person unicorn: Myth, miracle, or the future of startups?Read the passages on Polsia. The writer says that she doubts revenue claims of AI startups. Source for the outside law firm and infrastructure partner.
- C Rest of World, 29 April 2026, Viola Zhou: I got stood up by an AI agent, and tracked down its human owner in ChinaRead in full. A reporter found a Polsia customer after his agent sent her an email. Source for the customer case and the founder's one-sixth figure.
- C GTMnow podcast, 1 June 2026: Inside the Company that Raised $30M at a $250M Valuation With 0 EmployeesRead the transcript sections on pricing, costs, models and the fundraising. Source for the cost per task, the cheaper models and the period in Paris.
- C Pulse 2.0, 25 May 2026: Polsia: $30 Million At $250 Million Valuation Raised For AI Operations PlatformRead in full. It reports the founder's LinkedIn post. Source for the investor list and the supplier list.
- C RuntimeWire, 29 July 2026: Polsia's Ben Broca runs 10,000-customer AI business with zero employeesRead the free part. It notes that the customer and revenue figures come from the founder, not from audited statements.
- C Polsia Terms of Service and app code, read 12 October 2026The page loads by script, so the text was read in the app's JavaScript file. Source for the 3% and 20% fees, the 31-day hold, the credit steps and the missing live page.
- A TrustMRR profile of ad-flow.polsia.app, a customer company hosted on Polsia, not Polsia itselfRead the markdown profile. Stripe-verified revenue of US$0.00, data as of 12 October 2026. TrustMRR has no profile for Polsia itself, so this source verifies none of Polsia's own figures.
- C andrew.ooo, 4 March 2026, updated 8 October 2026: Polsia: How One Solo Founder Hit $1.5M ARR in 30 Days With Zero EmployeesRead in full. A publication that AI agents write under human supervision. Used to find other sources, not for numbers.
- C Loot Drop database export, 1,749 failed startupsLocal export. AI-written summaries, not independently verified. Source for the failed predecessors.