REA turned cloning software into a prompt
For forty years, the moat around paid software was the compiled binary. You could buy the product. You could not read it. That moat is draining. REA, short for Reverse Engineer Anything, is an open-source tool that connects coding agents to the disassemblers professionals use, Hopper, Ghidra and IDA. It lets an agent inspect an application without its source code. Its own README makes the promise plainly: an agent can explain how a feature works and "build a version for your project."
The response was not niche. On October 12, the GitHub API counted 149,178 stars on the repository, which was created in April and ships under the MIT license. Around it, a second wave arrived. KDnuggets reports agent-built clones of Adobe and Microsoft Office apps. The developer of WordCraft estimates about 62% real-world feature parity with Word. DeckCraft reports about 55% readiness as a PowerPoint replacement. Testers found bugs and missing features. The KDnuggets author is blunt: "REA cannot automatically produce a perfect clone."
It does not need to be perfect. A clone that covers the features a team actually uses is enough to cancel a subscription. That is the real event here. It is not a legal story yet. No vendor has been reported to sue over REA or the clones. It is a fiscal story, because every cancelled license takes a line of taxable revenue with it, and the people who wrote that software were on a payroll.
Citrini wrote this chapter in February
In February 2026, Citrini Research published The 2028 Global Intelligence Crisis, a memo written from the future. The authors call it "a scenario and not a prediction," and that label matters. In their fiction, firms watch internal teams replicate "six-figure SaaS contracts in weeks." The threat alone changes pricing negotiations. Software margins compress. Payroll follows. Eight months later, REA and the clones look like the first page of that chapter, arriving early.
The scenario's sharpest line is about tax. Citrini writes that the federal revenue base is "essentially a tax on human time." Payroll taxes and income taxes ride on hours worked and wages paid. In the scenario, labor's share of output falls to 46% and federal receipts run 12% below CBO projections. Those numbers are fiction. The mechanism is not. If output grows while hours and wages shrink, a revenue system built on human time collects less from a richer economy.
Citrini's name for that output is Ghost GDP: value that "never circulates through the real economy." The critics deserve a hearing. Citadel Securities answered within days, citing data that showed demand for software engineers up 11% year over year in early 2026. Its strategist argued that if compute costs more than labor, "substitution will not occur." Fair. But a clone is built once. After that, the compute bill is small and the license fee is gone for good.
A free clone removes two tax bases at once
Follow one subscription. A company pays a vendor for a design tool every year. That payment is revenue to the vendor, which pays engineers, support staff and sales teams. Their wages carry payroll contributions. The vendor's profit carries corporate tax. Now the company points an agent at the tool, builds a clone that covers its own workflows, and cancels. The license revenue disappears. The payroll that served it shrinks. Both bases thin out at once.
The value has not disappeared. The company still edits its images. It now does so without paying anyone. The saving lands as lower costs, which means a fatter margin on the company's own books. Some of that margin will surface as profit and pay corporate tax. Much of it will not. It can be reinvested, deducted or booked wherever profit is cheapest to book. The social contributions that used to travel with the vendor's wages do not come back at all.
This is Ghost GDP turned inside out. Citrini's version is output that is measured but not shared. The clone version is value that may not even be measured, because a tool with no price leaves no sale in the accounts. That is the fiscal risk of the open-sourcing wave. It is not that software becomes free. It is that the value software creates stops passing through any line a tax system can read.
This is the hardest test for our own calculator
Tax The Agent proposes an agent contribution: a rate on the revenue a firm's autonomous systems materially create or protect. It is a campaign proposal, not law anywhere. Software clones are the toughest case for it, and the campaign should say so first. A clone has no revenue. An MIT-licensed project given away on GitHub has no margin at all. If the base were the price of the agent's product, the base would be zero.
The base is not the product's price. It is the deploying firm's margin. The calculator's first field reads "revenue your autonomous systems materially create or protect." A cancelled license is protected revenue in plain sight. The firm that used to send a fee to a vendor, and through it to a payroll, now keeps that fee because an agent built and maintains its replacement. That retained amount is attributable, documented by the old invoices, and auditable.
The consequence is clean. The open-source developer who publishes a clone owes nothing, because they capture nothing. The hobbyist who reverse engineers a game to learn owes nothing. The firm that replaces a seven-figure software contract and the team behind it with an agent-maintained copy owes a contribution on the margin it now keeps. Free software is not the problem. Free margin is.
Why not tax the inference, as Citrini's scenario does
Citrini's fictional government reaches for two tools. A "Transition Economy Act" funds transfers partly through "a proposed tax on AI inference compute." A "Shared AI Prosperity Act" creates something between a sovereign wealth fund and "a royalty on AI-generated output." The second is close to this campaign's position. The first shows why compute is the wrong meter, and software clones make the case better than any example yet.
A clone is a burst of computation followed by years of near silence. The agent spends heavily for a few weeks to read the binary and rebuild the features. Then the copy runs like any other program. An inference tax collects during the build and then stops, while the cancelled license and the shrunken payroll stay gone. It also charges the student studying a binary at the same rate as the firm that replaces a vendor. Compute tells you how hard the machine worked. Margin tells you who kept the value.
Run the clone test on your own stack
This debate will not wait for a perfect clone. Pick one organization you know, your employer or a client. List its five largest software subscriptions. For each one, ask a blunt question: could an agent with a tool like REA rebuild the part we actually use within a quarter? Add up the fees for the ones where the answer is yes. That figure is margin about to move.
Then run it through the illustrative calculator below as agent-attributable revenue, and set the rate you would defend in public. Ask what share of those fees used to become wages at the vendor. That is the payroll exposure. You will hold a number the current tax code does not see. That is the whole case for an agent contribution in one spreadsheet row.
