Instinct is worth $10 billion. You still can’t sign up for it.
Those two facts sat next to each other in my notes all week and I couldn’t shake them. On Monday the company confirmed a $1 billion Series C at a $10 billion valuation, with Sequoia Capital, Benchmark Capital, and Coatue on the cap table. The product? A personal agent that does things on your behalf — a concierge that handles bookings and makes phone calls. And it’s still in invite-only early access, where it’s been since launching in August 2026.
I build bots for a living. I’ve shipped agents that call APIs, agents that read calendars, agents that chain six tool calls together and fall over on the seventh. So when a company gets valued at four times its previous mark for shipping an invite-only concierge, my first reaction isn’t skepticism about the valuation. It’s curiosity about what they solved that the rest of us haven’t.
The phone call is the hard part
Bookings and phone calls sound modest. They’re not. If you’ve ever wired up an agent to complete a real-world transaction, you know where the pain lives, and it isn’t in the language model.
It’s in the parts nobody demos:
- State that survives interruption. A restaurant puts your agent on hold for four minutes. Does it still know what it wanted? Does it know what it already said?
- Turn-taking with a human who doesn’t cooperate. Real people interrupt, mumble, transfer you, and ask questions your prompt never anticipated.
- Irreversible actions. A booking is a commitment. There’s no dry-run mode for a reservation under someone’s real name.
- Failure that degrades gracefully. When the agent can’t finish, does it hand back to the user with a clear account of what happened, or does it silently do nothing?
That last one is the tell. Most agent demos I see optimize for the happy path. Production agents live almost entirely outside it. A concierge that works has to be honest about what it couldn’t do, and that honesty has to be engineered — it’s error handling, retry logic, and escalation paths, not a personality trait you prompt into existence.
Why invite-only is a design decision, not a bottleneck
The instinct (sorry) among builders is to read closed access as a sign the thing isn’t ready. Sometimes that’s true. But for agents that take irreversible actions in the physical world, a small user base isn’t a limitation — it’s the only sane way to operate.
When your bot books a table, the failure mode isn’t a bad response the user ignores. It’s a restaurant holding a reservation for someone who isn’t coming. Every wrong action has a cost borne by a third party who never agreed to test your software. Scaling that before your success rate is high enough is how you end up with businesses blocking your outbound calls.
If you’re building in this space, steal the pattern. Gate access. Watch every transcript. Fix the specific failures you see instead of the ones you imagine. It’s slower and it’s correct.
What $10 billion is actually pricing
The jump from $2.5 billion to $10 billion in one round is the number people will argue about. I read it less as a bet on the concierge feature and more as a bet on the plumbing underneath it.
A personal agent that handles bookings is a narrow product. But the infrastructure required — persistent memory across sessions, reliable tool execution, real-time voice, permission models for acting on someone’s behalf, graceful failure — is general. Solve it for restaurants and you’ve solved most of it for appointments, returns, rescheduling, and every other errand that currently requires a human on a phone.
That’s the thesis worth paying attention to as a builder. The demo is the concierge. The asset is everything that makes the concierge not embarrass you.
What I’d take back to my own code
You don’t need a billion dollars to apply the lessons here. A few things I’m changing in my own agent projects after sitting with this news:
- Treat every external action as a transaction. Log the intent, the attempt, and the outcome separately. You cannot debug an agent from its final output alone.
- Build the escalation path before the happy path. Decide what “I couldn’t do this” looks like on day one, not after your first angry user.
- Narrow the scope until reliability is boring. One task done at a high success rate beats ten tasks done at a coin flip.
- Keep humans in the confirmation loop for anything irreversible. Autonomy is a spectrum, not a switch.
Instinct raised a billion dollars on a product most people still can’t touch. That gap between valuation and availability is the whole story, and it points at something useful: in agents, the money is following reliability, not reach. The teams that win won’t be the ones with the longest feature list. They’ll be the ones whose bots can be trusted to pick up the phone.
đź•’ Published: