Why Instinct Is Worth as Much as Snap

Personal agents replace the interface with intent. Investors are paying for what that does to the customer relationship. Also available as a video.

A lot of the conversation in AI right now is converging on personal agents. Meta shipped Muse in early September. OpenAI followed with Dots three weeks later. And around the same time, Noah Shinn gave his first long interview about Instinct.

Instinct is a tiny company. It's invite-only, it has no app, and it reportedly has just over 100,000 users. On September 28 it raised $1 billion at a $10 billion valuation, about a month after raising at $2.5 billion.

Compare that to Snap. Snap has hundreds of millions of daily users, billions in annual revenue, mature ad infrastructure, and a product that's already global. Its market cap is around $9 billion. So a startup in early access is priced in the same neighborhood as a public company with several thousand times as many users.

On current fundamentals, that comparison makes no sense. So why were investors willing to pay it, and what are they trying to get ahead of?

A price on future possibility

Start with who's making the call. Snap's valuation gets tested every day by a liquid public market. Instinct's comes from a single financing round negotiated among a small set of venture investors.

That doesn't make the number fake. It does make it much more explicitly a price on what the company could become. And what it could become depends on whether the basic model of how we use software changes.

From interfaces to intent

The surface-level description of Instinct is an AI assistant that takes actions. It books travel, makes reservations, cancels subscriptions, reschedules appointments. People mock the demos for exactly this, and fair enough. Booking a flight is a toy example. It's in every demo because it's easy to reason about and quick to show.

The more important shift is from UI-mediated software to intent-mediated software. Shinn's own pitch is that "there are no new interfaces." You text it or call it.

Today, the user opens an app, the app has an interface, and the interface calls an API or a service. We do almost all of the procedural work ourselves. We learn the menus, compare options, fill out forms, move between tabs, and keep track of what's in the cart. The software exposes capabilities, and it's on us to assemble them into an outcome.

In the agent model, the user states an objective. A planner then decides which tools, APIs, and browser actions to use, and runs the loop end to end. The interface layer between the human and the internet gets replaced with something new.

From commands to standing mandates

"Book me a flight" is still a command. The agent does one task and stops. The bigger change comes when you stop thinking in commands and start thinking in policy: "Never let me miss a meeting because of travel."

Imagine a busy parent hauling kids to practices, games, and appointments. The command version is "schedule my kids' appointments this week." The policy version is "keep my kids' schedules coordinated so they always have a ride and nothing important overlaps, and tell me when there's a conflict I need to deal with."

That's a standing mandate.

Once software can hold a mandate over time, the question becomes which resources you're letting it control. Your money? Your calendar? Your communication? Your recurring obligations? At that point the core product isn't the chat interface. It's a permissioned system for turning high-level intent into bounded action.

That comes with a different set of technical problems. What is the agent authorized to do? How long does that authority last, and under what conditions? Can the agent delegate? What happens when a mandate is violated, and who's liable? How do you prove after the fact that an action was authorized?

None of those are UX questions. They're identity, security, auditability, and governance questions. And they matter because an agent is only useful if you trust it enough to stop watching it act.

The moat is a learned model of you

There's a running joke that every time OpenAI or Anthropic ships a new model, someone announces that a hundred startups just lost their moat. So what would Instinct's be?

The easy answer is data: it knows your calendar, your preferences, your purchase history. But raw data probably isn't the deepest moat. A new agent can import your calendar and your Amazon orders on day one.

What it can't import is the policy function the old agent learned for you. Given this context, these options, and these constraints, what would you choose? Maybe you usually take the cheaper hotel, except when the expensive one saves you 40 minutes. Maybe you care about price on personal travel and location on work travel.

That isn't a database of facts. It's a learned model of how you allocate money, time, risk, and convenience, and it's hard to reproduce. Those learned decision boundaries are the switching cost. If your software gets better the longer you use it, you won't want to start over with something that suddenly doesn't know you.

Agents attack inertia

That same aversion to switching props up a huge amount of the current economy. People don't shop for new insurance every month. They don't renegotiate subscriptions or move banks every time a better savings rate appears. Human attention is scarce, so businesses earn rent from inertia, switching costs, information asymmetry, and the fact that nobody can optimize everything all the time.

A persistent agent with your context can. It can search, compare, renegotiate, and switch continuously. "Find me cheaper insurance" becomes "keep my household insured at the best available price, subject to these coverage constraints."

Software like that could shrink entire categories of friction that businesses have depended on for a long time. If enough consumers use it, margins compress across industries.

It also creates buyer power. Say an agent layer, Instinct's or one Amazon builds, eventually routes tens of billions of dollars a year in purchasing intent. It can go to merchants and say: we control a large pool of demand, so give our users better pricing, better inventory, better cancellation rights. More users means more aggregated demand, which means better terms, which means better outcomes, which means more users. That's a stronger network effect than "people use the product because it's good." And it's worth far more than a take rate on each transaction.

Whose side is the agent on?

Now the problems. If the agent is supposed to represent the user, but its business model depends on merchant take rates, who does it work for?

Say it's choosing between two hotels. Hotel A is better for you. Hotel B pays the agent more. Which one gets recommended?

Search and marketplaces have always had this conflict. Google started as a neutral search layer and built an ad business on top of it. Amazon built sponsored placement into its rankings. The personal agent has the same problem with a higher trust requirement, because the whole point of the product is that you don't inspect every decision.

We may end up with two categories: a fiduciary agent that works only for the user and charges them directly, and a commercial agent that's free and paid by merchants.

Ads, websites, and the exception interface

It's tempting to say agents kill advertising. More likely, advertising changes form. If the user is no longer browsing pages, advertisers stop competing for human attention and start competing for machine routing. The ad becomes structured economic data that an agent can evaluate directly: price, availability, commission. Persuasion turns into machine-readable bidding. Humans see fewer ads while a lot more machine-to-machine negotiation happens underneath.

Websites probably split the same way. Businesses maintain two interfaces. The human-facing one handles branding, storytelling, and discovery. The machine-facing one exposes deterministic APIs for availability, pricing, transactions, identity, returns, and permissions. The API becomes the primary economic interface, and the website becomes the secondary one.

The app changes too. If the agent does most of the work, the ideal interface isn't a dashboard showing everything that's happening. It's an exception interface:

  • Flight changed. Rebooked automatically.
  • Insurance renewal up 23%. Two alternatives found. Approval required.
  • Unusual purchase blocked.
  • Dinner moved to 8. No action needed.

The human only sees the cases where judgment or authorization is needed as a final pass.

Being right about the future isn't owning it

Even if this worldview is right, it doesn't follow that Instinct captures the value.

Agent interoperability could get standardized. Permissions could become portable. Payments and identity could stay with other platforms. If Google, Apple, Visa, Mastercard, OpenAI, or some future protocol layer controls the underlying infrastructure, Instinct could be correct about where things are going and still fail to own the outcome.

And incumbents won't help. Amazon has no interest in handing its customer relationships to an outside agent that bypasses its sponsored listings.

People make this mistake a lot: being right about a technology and assuming that makes them right about a particular company. The web became enormous, and Netscape didn't capture it. Mobile became enormous, and most early mobile companies didn't win.

So the $10 billion valuation is two bets at once. The first is that delegated agents become a major computing layer. The second is that Instinct becomes one of the companies that controls it. I think the first one happens in some form. The second is a toss-up.

Venture capital is fine with that. Investors don't need every bet to work, and if this one does, the upside is enormous. The winner will be the company or protocol that holds persistent authority over the user's intent, money, and time: the one that owns the permissions, the policy model, the demand, and the point where every other business has to negotiate for access to the customer.

That's the same thing I argued last time. When AI makes creation cheap, access is what gets expensive.

The internet as we know it was built around capturing attention. Incumbents spend enormous amounts of money getting you to open their apps and stay there. If personal agents get good, businesses that make money by controlling what you see have a serious problem. Businesses that provide the underlying service don't. You still need the thing you're buying.

That's why 100,000 users doesn't bother the investors. They're betting that opening apps, navigating interfaces, and filling out forms will someday feel as dated as getting a call on a landline.

Sources