What happens to a labor argument when a growing share of the labor isn’t human?
On October 8, 2026, the Trump administration announced it is suspending Microsoft from the program used to apply for green cards on behalf of workers who come to the U.S. on H-1B visas, accusing the company of fraud. Adobe and other companies are included in the suspension. Vice President JD Vance framed it in arithmetic: “If you do the math, for every worker that Microsoft laid off, they replaced that worker with one and a half foreign indentured servants.”
I work on agent architecture, not immigration law, so I want to be careful about what I can and cannot claim here. The fraud allegation is an allegation. I have no independent visibility into Microsoft’s filings. What I can speak to is the thing that makes this particular fight different from the H-1B fights of the last two decades, and it has nothing to do with visas.
Headcount stopped being a clean proxy
The structure of Vance’s claim is a substitution argument: a slot opened, a slot closed, and the ratio tells you something about intent. That logic works when a job is a stable unit. One engineer leaves, one engineer arrives, and you can reason about who got displaced and why.
That unit is less stable than it was. Inside large software organizations, the work that used to define a mid-level engineering role is being decomposed into pieces that sit at different places on the automation curve. Code generation, test scaffolding, log triage, dependency bumps, first-pass code review, documentation upkeep, on-call summarization — these are now routinely handled or pre-chewed by agent systems. What remains in the human role shifts toward specification, judgment, system decomposition, and the review of machine output.
The headcount number does not record that shift. A company can reduce roles in one function, add roles in another, and route a third category of work to inference capacity that shows up on the balance sheet as compute spend rather than payroll. Any ratio drawn across those categories is comparing things that aren’t the same thing.
Why this matters for how agents get built
This is where the policy question stops being abstract for anyone working on agent systems.
Agent engineering is still an apprenticeship discipline. There is no textbook that tells you how to set a tool-call budget, where to put a verification step, how to decide when an agent should escalate to a human, or how to debug a failure mode that only appears after nine steps of accumulated context drift. That knowledge lives in the heads of people who have shipped these systems and watched them break in production. It transfers by proximity.
Immigration policy is one of the inputs that determines where that proximity happens. The suspension targets the green card pathway specifically — not the visa itself, but the route from temporary status to permanence. That distinction matters for a field built on long-horizon tacit knowledge. Someone on a three-year clock optimizes differently than someone building a decade-long career. Deep architectural work, the kind that requires sitting with a hard failure mode for eighteen months, favors stability.
I’ll say plainly what I don’t know: whether this suspension is temporary, how many filings it affects, or what the companies will do in response. Those details aren’t available in what’s been reported, and I’m not going to guess at them.
The measurement problem nobody is solving
There is a version of the displacement debate worth having. Automation does change labor demand. Workers do get displaced. Those are real and the people affected deserve better than slogans from any direction.
But having that debate well requires measurement that doesn’t currently exist. We have no shared accounting for how much work an agent system absorbs, no standard for distinguishing augmentation from replacement, no way to separate a layoff driven by automation from one driven by a reorg or a product sunset or a bad quarter. Companies don’t publish it. Regulators don’t ask for it in a form that would be meaningful. So the argument gets conducted with headcount, because headcount is the only number everyone can see.
That’s a weak foundation for policy with this much blast radius. Visa and green card decisions shape research careers over decades. Building them on a ratio that mixes human roles with work that may have been automated produces conclusions you can’t audit.
The useful demand, for whichever administration is making these calls, is better instrumentation. What work moved to machines, what work moved to people, and where did those people need to be standing to do it. Until someone can answer that, the arithmetic stays rhetorical.
đź•’ Published: