Product1 publisher3 min readPublished
Instinct's agent proposes the errand before the user has to think of one
Instinct, an AI agent people text with, is reportedly in talks to raise $1 billion at a $10 billion valuation, according to The Information. A Wired tester credits its habit of proposing errands, but the public numbers so far count installs and investor money, with retention unreported.
The Product Desk · Product desk
What happened
- A writer who had just left a full-time job at Wired gave Claude Cowork email and calendar access and found it behaved like a regular chatbot with a few extra tools.
- Instinct is an invite-only agent that works over iMessage and WhatsApp, connects to email, calendar and messaging apps, and has been in private beta since February.
- Given a Venice work itinerary, Instinct booked the writer's restaurant reservations, sending WhatsApp messages to small local spots to do it.
- The writer deleted a rival agent, Lindy, after one week because it texted every morning and joined Zoom meetings without being asked.
- Meta's Muse assistant, launched this month, is the top free app in Apple's App Store with more than 900,000 downloads by third-party estimates.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint A download estimate or a reported valuation tells an operator nothing about retention or repeat use, so the business case for rolling out one of these agents still needs usage data of its own.
- decision Agent builders now have two settings to choose: who proposes the task, and how often the agent speaks unprompted. Lindy's quick deletion shows what happens when the second is set too high.
- exposure Once users hand consumer agents their inboxes and calendars, a single flaw like Muse's Mac vulnerability can reach far more personal data than a chatbot bug could.
- contradiction Muse is described as calling restaurants for its users, yet 404 Media reports that people in call centers make some of those calls, so a completed booking says little about what the software itself can do.
"What did I want it to do? I had no idea," the Wired tester wrote of the first stretch with Claude Cowork [2]. The best earlier result from Claude had been a set of freelance invoices, about 10 minutes saved [9]. Jasmine Sun, a technology journalist, wrote that "most people's problems are not software-shaped, and most won't notice even when they are" [8].
Teams building agents tend to picture a user who arrives with a job. This user arrived with inbox and calendar access and an expectation of a "hyper-capable digital assistant" [1], and no job in mind [2]. Instinct's answer is to remove the open text box. Users get iMessage, WhatsApp and a few prompts, and the agent suggests things it can do before doing them [10]. According to the author, AI researchers who talk about Instinct use the phrase "form factor" a lot [11].
Restraint counts as much as suggestion. Instinct "doesn't text too much," the tester wrote [12]. Lindy messaged every morning and joined Zoom calls unasked, and it was deleted within a week [13].
The case made for agents, as the author describes it, is automating "the majority of people's administrative drudgery" [18]. The first thing that worked here was booking restaurants for a Venice trip [14]. The tester called reservations "a gateway drug for many agent-curious people" [15]. For Muse, even that task comes with a caveat. 404 Media reported that some of the restaurant calls Muse places are made by humans in call centers [16]. Muse also shipped with a flaw that Ars Technica said would have "let attackers do 'whatever' they wanted on a victim's Mac" [7]. The author is "not about to give Meta a ton of my personal information" [17].
If Instinct's round closes on the reported terms, the company will have raised $1.35 billion [5]. Muse's 900,000-plus figure is a third-party estimate of downloads [6]. The case that these agents are reaching ordinary users comes down to one tester's account and those two figures.
For a team choosing one of these tools, I'd sort them on two axes. One is who names the task, the user or the agent. The other is whether the agent waits or reaches out on its own schedule. Cowork, as the tester used it, left the naming to the user and waited [1]. Lindy acted on its own and reached out daily [13]. Instinct suggests tasks and texts sparingly [10][12]. In my view, the useful test for a product in Instinct's box is the share of users who accept a second kind of task, beyond restaurant bookings, within their first two weeks. Neither a download estimate nor a valuation measures that share.
What to watch
- Whether Instinct's reported $1 billion round closes at the $10 billion valuation The Information described.
- Any retention or repeat-task figures from Instinct or Meta that go beyond third-party download estimates.
- Whether Instinct keeps its low-message, suggest-first design as it moves beyond invite-only access.