On 9 October 2026 Vercel published a short changelog entry: "Your agent can now use the Vercel CLI to buy a domain." An AI coding agent can search for candidate names, check availability and registrar pricing, and start the purchase, all from the command line.
The interesting sentence comes next: "The purchase decision itself stays with you by design."
How the human step is built
The mechanism is specific. The agent uses vercel domains search to find candidates, vercel domains check and vercel domains price to check availability and pricing, and vercel domains buy to start the purchase. Then, in Vercel's words:
"Run non-interactively,
vercel domains buyreturns a structured, machine-readable error and the suggested next commands, so the agent hands the confirmation back to you, prohibiting it from spending money without your approval."
The person then confirms the price, the auto-renew choice and the registrant contact details, either in an interactive terminal or in the dashboard. Vercel also ships an agent-facing skill that covers the discovery, pricing and purchase workflow, so an agent follows the full sequence without being told the steps.
Two things are worth being precise about. This is one command and one capability: an agent buying a domain through the Vercel CLI. And "prohibiting it from spending money without your approval" is Vercel's description of its own tool. We have not tested it, and this article doesn't claim it can't be bypassed.
The same problem, solved the other way
A week earlier, on 7 October 2026, the QGIS project merged a change to its AI policy. Its earlier policy allowed AI tools on one condition: "there must be a human in the loop." Contributors had to review AI-generated code themselves before asking others to review it. The author of the change wrote that this "clearly didn't work". The replacement bans AI-assisted code contributions to the qgis/qgis repository outright. (We wrote about that change here.)
Putting the two side by side is our reading, not something either project said. But the contrast is useful. Both wanted a person to make the final call. QGIS's version depended on contributors remembering to do the review, and it gave up on that. Vercel's version doesn't ask the agent to remember anything. When no person is present, the purchase path simply stops and returns an error that names the next step. The human step is the only way through.
A question to ask about any agent you deploy
If you are commissioning software that includes an AI agent, the useful question is not whether a human is "in the loop". It is how the loop is enforced:
- What happens when no one is there? Does the action stop, or does it carry on and leave a note for later?
- Is the approval step the only code path, or a check that someone has to remember to run?
- What does the agent get back when it is stopped? A clear, structured answer it can hand to a person is what lets it stop gracefully instead of guessing.
- Which actions does it cover? Spending money, sending messages, deleting data, publishing. Each one deserves the same question.
A guardrail that depends on someone remembering to check is a hope. A guardrail that is the only path through the code is a design decision you can inspect.
Sources: Vercel changelog, "Agents can now buy domains with the Vercel CLI" (published 9 October 2026, read 14 October 2026); QGIS-Enhancement-Proposals pull request #387 (merged 7 October 2026).