Product1 publisher2 min readPublished
Instinct acts before its user can see what it remembers
An early hands-on with Instinct from the Ben's Bites newsletter finds the personal agent proactive and opaque, acting on meeting follow-ups while keeping what it stored out of the user's view. Muse is next, with reservations about Meta.
The Product Desk · Product desk
What happened
- Ben's Bites reported early access to Instinct, which the newsletter describes as the personal agent all the VCs are raving about, and rated it a bit meh after use.
- The newsletter said the product feels slow, that the magic is not there yet, and that getting these onboarding experiences right for everyone is really hard.
- Ben's Bites said these personal agents keep what they remember about the user, and the context they hold, out of the default view.
- The writer said he is waiting on Muse access but does not want to give Meta more information about him, calling the company an unreliable privacy partner.
- He expects every AI company to start shipping its own personal agent, the way every app added an AI assistant chat box.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint Whoever rolls one of these out has to answer for what the agent knows about colleagues and customers, and the default install keeps the memory hidden from the security reviewer who asks on Friday.
- decision Pilot owners now have to decide whether proactive action is on by default or opt-in, because the first unprompted message sets the user's guess about what the agent can already send.
- exposure Granting a personal agent mailbox and calendar scope puts the vendor's privacy record inside the trust chain, and the writer names Meta as the case where that gives him pause.
- constraint The only named comparison still calls its ratings provisional, so a buyer weighing Instinct against Muse is working from one first-person account, with the independent scores still unsettled.
Instinct's opening message to Ben, who writes the Ben's Bites newsletter, read: "I finished working on your meeting follow-ups - reply here and I'll send the details." [2] He said the line made him hope it would not start sending emails on his behalf, and he counts himself "pretty savvy (ish) with agents" [3]. The message reports work already done, and the details come only if he replies.
Proactivity is the feature that separates a personal agent from a chat box in most product decks. On the receiving end it arrives as an inbound request from something that has already acted. "Makes me feel like I'm having to do work to keep it happy, or like I've got a boss again - no thanks," Ben wrote [4].
What makes an unprompted action checkable is a page showing what the agent wrote down about you. He said he likes "being able to see and edit what's in my files to steer my agents" [6]. Where that page is missing, the user corrects a wrong assumption by chatting with the agent about something he cannot read.
This is one person's first hours with one product, and the newsletter published no retention, usage-depth or time-to-value figures [13]. The nearest comparison the newsletter points to is Assistant Benchmark, which tests everyday assistants including Grok bot, Muse and Instinct, and whose ratings the newsletter calls a work in progress [10]. Three products are named there [12].
Another item in the same issue asks whether an agent app should have bots or tasks and answers "Maybe neither. Because both make you organise the work." [11] That is the same complaint from the builder's side.
Two axes sort a pilot. Whether the agent acts before the user asks, and whether the user can read and edit what it has stored. The failure Ben describes sits in a single quadrant, proactive with hidden memory, because the first unprompted action lands before the user has any way to audit the assumptions behind it. In my view the ordering runs the other way: the memory view first, then the first unprompted message. Where the unprompted message is the demo, the week after it goes out gets spent explaining it.
What to watch
- Whether Instinct adds a default view of stored context and memory editing, and how onboarding changes once it does.
- Ben's Bites is waiting on Muse access; whether Meta exposes editable memory will test the same complaint against a different vendor.
- Whether Assistant Benchmark's ratings for Grok bot, Muse and Instinct firm up enough to check a single hands-on account against.