How to read this memo
What follows is a scenario, not a prediction. It borrows its form from Citrini Research's The 2028 Global Intelligence Crisis, a macro memo written from the future and published in February 2026, which its authors labeled "a scenario and not a prediction." Our memo is dated June 30, 2029. Its seeds are real and dated: REA, the reverse-engineering tool that passed 149,000 GitHub stars in October 2026, and Citrini's October 8 report asking "What happens when millions of agents continuously move deposits toward the highest bidder?" Everything after October 2026 is invented. The invented numbers are there to show a mechanism, not to measure anything. Where a figure is fiction, the memo says so.
June 2029: the invoice is gone
Memo, June 30, 2029. The strange thing about this crisis is that nothing was destroyed. Output in the sectors that stopped paying for software is higher than it was in 2026. Firms that once ran on forty subscriptions now run on agent-maintained code they own. Their margins widened. Their payrolls shrank. The value was all still there. It just stopped passing through anything with a price on it.
In the scenario's accounts, business software spending in the United States is a third below its 2026 level, a fictional figure. The money did not go elsewhere. It stayed with the firms that used to spend it, as cost savings that became margin. Some of that margin surfaced as taxable profit. Most did not. It was reinvested, deducted against the capital spending that the 2025 tax law made generous, or booked in the subsidiary where profit is cheapest. The payroll contributions that used to travel with the vendors' wages did not come back in any form.
Q4 2026: the clone wave looked like a hobby
The first quarter of this story was mistaken for a novelty. REA connected coding agents to the disassemblers professionals use, and its README promised that an agent could inspect an app without source code and "build a version for your project." KDnuggets covered agent-built clones of Office and Adobe apps with parity figures in the 55% to 62% range, and testers reported bugs. Vendors issued statements about intellectual property. No lawsuit of consequence followed, because the clones did not copy code. They copied behavior, and behavior had never been the thing the license protected.
The number that mattered was not parity. It was the cost of the first 60%. In the scenario, by spring 2027 a mid-sized insurer's internal team could replace a design tool, a ticketing system and a document platform with agent-maintained code in a single quarter, a fictional pace. Procurement departments added a line to every renewal negotiation: what would it cost us to have an agent build this? The answer was usually a number smaller than the renewal. Vendors cut prices to stay below it. Citrini's 2028 memo had predicted exactly this, firms replicating "six-figure SaaS contracts in weeks," with the threat alone reshaping pricing. It arrived a year early.
2027: build-not-buy becomes a clause
By the end of 2027 in the scenario, the standard enterprise contract carried a clause the lawyers called the clone option: the customer may terminate at any time on evidence that an in-house equivalent is in production. Vendors accepted it because the alternative was no contract. Software company revenue fell. Software company payroll followed, since the engineers, support staff and sales teams that revenue had funded were the first cost cut. Those wages had carried payroll contributions. Those contributions stopped.
The frontier labs did well. Enterprises had already been moving toward closed models, Menlo Ventures had measured open-source share of enterprise usage falling from 19% to 11% by late 2025, and the clone wave made closed intelligence more valuable, not less. The firm that bought the best closed model and used it to clone every open application kept the margin from both ends. Tax authorities noticed a pattern they could describe but not tax: AI revenue was concentrating in a few labs, which paid corporate tax on it, while the far larger savings were spreading across every deploying firm as cost reductions that no base was designed to see.
2028: agents pay agents on rails that never close
The second seed sprouted in 2028. Citrini's October 2026 report had argued that "AI agents move programmatically, 24/7, across applications" and asked whether an agentic bank run was coming. In the scenario, the bank run came in a mild form: deposits chased yield in minutes, and banks paid up for stable funding. The larger shift was quieter. Agents started buying from agents. A procurement agent at one firm settled directly with a fulfillment agent at another, on tokenized rails that existing finance, "gate-kept by layers of intermediaries, legal authorities and KYC processes," could not match for speed.
The first-order effect was efficiency. The second-order effect was fiscal. An invoice is a tax document. It names a seller, a buyer, a price and a date, and it anchors sales tax, income attribution and transfer pricing. When agents settled continuous micro-transactions on programmable rails, the invoice did not vanish, it dissolved into a stream. Tax authorities found they could see every transaction and attribute none of them. In the scenario, a fictional interagency memo described the problem as perfect visibility with zero attribution. The ledger showed the money. It did not show whose work the money paid for.
2029: the first agent contribution, and the fight over attribution
In the scenario, the first jurisdiction to enact an agent contribution did so in early 2029, as a reporting requirement first and a rate second. Firms above a revenue threshold had to report attributable autonomous revenue, the share of revenue their autonomous systems materially created or protected, using documentation standards borrowed from transfer pricing. The rate began at zero. The published aggregate was the number politics had been missing. The rate moved the following year, tied to a measured decline in the jurisdiction's payroll base.
The fight was over attribution, as it was always going to be. Firms argued that an agent-maintained clone was simply software, like the spreadsheet before it. The authority answered with the old invoices: this is what you paid, this is the payroll that fee supported, this is the revenue the replacement now carries. The open-source authors who had published the clones were never in the room. They kept no margin and owed nothing. The agent-to-agent settlement streams, which had looked like the end of attribution, became its best evidence. Every stream carried an agent identity tied to an accountable principal, because without one no counterparty would settle. Identify, attribute, return. The rails built the first step on their own.
What the scenario is for
Back to 2026. The memo is fiction, and its numbers are placeholders. Its mechanism is not. Two things are true today. Agents can read compiled software and rebuild its behavior, which removes the price from the application layer. Agents are beginning to move money programmatically on rails that do not close, which removes the invoice from the transaction layer. A tax system built on prices and invoices loses both at once, while the value it was meant to share grows.
The campaign's proposal is the same in fiction and in fact: an agent contribution on the revenue a deploying firm's autonomous systems materially create or protect. It is not law anywhere. Its first step costs nothing, a reporting line at a zero rate. One exercise for this week: pick a firm you know, estimate the share of its software spend an agent could replace within a year, and run that figure through the illustrative calculator below as protected revenue. That is the margin the memo is about, and it is already moving.
