- The Verge tested the AI agent Instinct. It got swim lesson times right but missed a prerequisite class, and the reviewer caught it only by checking the page.
- Agents that act on your behalf can leave out a condition without any sign. The Verge does not say why this miss happened, and it was a single test.
- Leaders should ask what accounts an agent can reach, who verifies its results, and how it is paid for.
A personal AI agent is a new kind of reader. It reads the web for you, then acts on what it finds. The Verge's hands-on test of a startup agent called Instinct shows one risk of that arrangement: a result can look complete and still leave out a rule that matters. It also raises a second question for anyone approving these tools. Who pays for the agent, and who else gets a say in what it does?
What the reviewer found
The Verge tested Instinct against two rivals, Meta's Muse and OpenAI's Dots. Instinct launched in August as an invite-only product with almost no marketing. The reviewer says it held its own against both larger companies' agents.
The tasks were everyday errands. Instinct emailed an eye doctor's office, started an Ikea return, and worked out which Ikea product the reviewer meant from a loose description. It also flagged a possible calendar conflict that nobody asked it to check.
One task went wrong. The reviewer asked Instinct to find swim lessons for a preschool-age child and pointed it to a community center. The times and dates were correct. But the center requires a prerequisite class for preschool-age lessons, and Instinct missed it. The reviewer caught the gap only by following along in a browser.
How these agents work
The Verge describes these agents as consumer versions of tools that software developers have used for the past year, such as Codex and OpenClaw. Three parts matter.
First, the interface. You talk to Instinct by text message, through iMessage, WhatsApp or email. There is no app. The agent sits wherever you can send a text.
Second, the access. The agents have a virtual machine, which is a computer running in the cloud that they can use to browse and fill in forms. They can also connect to your personal email and calendar. Instinct links to services including Google Workspace, Slack and Notion. When it needs a login, it sends a link to a "secure vault" where you enter your details.
Third, the reading. The agent works through ordinary web pages, which people made for people. The Verge does not say why Instinct missed the prerequisite, so the cause is unknown. The reviewer ties the miss to a wider problem: real web pages are often untidy, and a vital detail may sit inside a pop-up window. The article does not say the swim class rule was in one.
The idea: a tidy answer can hide what was skipped
Here is the lesson. Whatever caused this miss, the result had a particular shape. The agent returned correct dates and times. Nothing in them signaled that a condition was absent. A person who only reads the agent's reply has no cue to look further.
This test involved one agent, one reviewer and one missed detail. It does not show how often agents fail. The reviewer called the correct times and dates "one very low hurdle cleared." The condition that mattered was the part left out.
For any organization that publishes prices, eligibility rules or booking terms, the point is practical. If agents become a common way for visitors to read your pages, a rule that is hard to find may be missed more easily. That is a plausible risk, not something this test proved.
Who pays for the agent
Instinct is free for invite-only users, and the founder, Noah Shinn, says the intent is to keep it "as affordable as possible." Meta's Muse is free. OpenAI's Dots is for paid subscribers only.
Analyst Avi Greengart of Techsponential told The Verge that free products may carry hidden costs. "If you're not paying for it directly, then you're paying for it indirectly," he said. He also suggested that goods and services may cut deals to get access to your agent.
Greg Ireland, a senior director at IDC, expects advertising to play a role. The Verge notes that Instinct has already tried pushing "recommended products" to users.
That is analysts' expectation, not a finding about any agent's behavior. But it names a fair question. An agent that books, buys and recommends on your behalf is a new place where commercial influence can enter.
What leaders should ask
Start with access. Which email, calendar, chat and document accounts can an agent reach, and who approved that? Instinct's own setup shows how short the path is from a text message to a Google Workspace or Slack account.
Then ask about checking. For tasks with a cost or a deadline, does a person verify the result against the source page? The reviewer's catch came from doing exactly that.
Ask about incentives. How is the agent paid for, and does its maker earn money from the products it suggests? The answers belong in any procurement review.
Finally, look at your own pages. The Verge does not say a pop-up caused this miss. Still, putting a key condition in plain page text is a low-cost way to make it easier for any reader, human or machine, to find.
An agent can do the errand, but someone still has to read the fine print.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: The Verge.
