Instinct AI: What a $10B Personal Agent Teaches Builders

Instinct AI is a personal agent you text or call, and it hit a $10 billion valuation less than a year after it was founded. Its 23-year-old founder, Noah Shinn, co-wrote Reflexion, a well-known paper on agents that learn from their own mistakes. For builders, Instinct offers two lessons. Agents that act for people can grow fast, and they need strong safeguards from day one.

Key Takeaways

• Instinct raised $1 billion at a $10 billion valuation on September 28, 2026, a month after raising $250 million at $2.5 billion.
• It's a personal agent you text or call. It plans trips, orders groceries, books tickets and cancels subscriptions, using its own phone and computer.
• Founder Noah Shinn, 23, was lead author of Reflexion, a 2023 paper on agents that learn from their mistakes, and an early employee at Sierra.
• Some early users reported unexpected actions. Instinct has since added sandboxes, short-lived credentials and a hallucination check before it acts.
• For builders, the lesson is simple: reflection and memory make agents better, but permissions and isolation make them trustworthy.

What Is Instinct AI?

A personal assistant that does tasks, not just answers questions.

Company description: Instinct says its assistant "understands what you're working on and what's important to you." It adds: "It's trained to use a phone and a computer. You can text or call it. It uses its devices in the same way that humans do."

In practice, people text it requests like these:

• Plan a trip and book the flights and hotels.
• Order groceries or other everyday purchases.
• Book tickets and restaurant tables.
• Cancel subscriptions, which some users say has saved them hundreds of dollars.

A newer concierge feature calls businesses that don't take online bookings, TechCrunch reported. A "trusted person network" lets two users' agents coordinate plans with each other. Shinn wants the assistant to learn each user's preferences over time, so people don't have to explain the same things twice.

Instinct is still invitation-only. Demand is high enough that some users have tried to sell their invitations on eBay, VnExpress reported.

How Does Instinct AI Work Behind the Scenes?

Instinct hasn't published a technical paper on its product, but its announcements and press coverage describe the main parts:

• Its own phone and computer. The agent uses a phone number and a computer, and operates apps and websites the way a person would.
• Text and voice. Users reach it by text message or by phone call, with no new app to learn.
• Access to your accounts. The company says it can work with email, messages, screen, audio and location, with the user's permission.
• Phone calls to businesses. The concierge feature calls places that don't offer online booking.
• Memory of preferences. It's meant to learn how you like things done, so you don't repeat yourself.
• Agent-to-agent plans. Two Instinct users' agents can coordinate with each other.

That last point is new. Once agents talk to other agents, mistakes can spread between users, so isolation and clear permissions matter even more.

How Fast Has Instinct AI Grown?

Faster than almost any consumer software company on record.

When	Milestone
Late 2025	Shinn leaves Sierra and founds Instinct, per his LinkedIn profile
Early investment	Valued at about $50 million
February 2026	Private beta begins
August 2026	Invite-only launch; $250 million raised at a $2.5 billion valuation
September 28, 2026	$1 billion raised at a $10 billion valuation

The September round came from Sequoia Capital, Benchmark and Coatue, Reuters and TechCrunch reported. That's about 200 times the early valuation, within roughly five months.

Company claims: On the "Invest Like the Best" podcast, Shinn said Instinct was approaching $1 billion in yearly transaction volume, with travel about half of it. "We spent $0 on marketing so far," he said, as quoted by Fortune. These figures come from Instinct and haven't been independently checked.

Much of the early spread came from investors themselves. "Everyone is just hounding each other for access," Trace Cohen of Six Point Ventures told Business Insider.

Who Is Noah Shinn, and What Is Reflexion?

A researcher who wrote one of the best-known agent papers while still an undergraduate.

Weeks after ChatGPT launched in late 2022, Shinn was a student at Northeastern University. He answered a LinkedIn post from Ashwin Gopinath, then at MIT, who wanted students to try the new technology. Over about three months, they and other researchers built Reflexion, which was presented at the NeurIPS conference in 2023.

The idea is easy to explain:

• Try the task. The agent attempts a job, such as writing code.
• Check the result. A test or a score says whether it worked.
• Reflect in words. If it failed, the agent writes a short note on what went wrong and what to try next.
• Remember and retry. The note goes into memory, and the agent uses it next time.

No retraining of the model is needed. The agent improves by learning from its own written feedback. In the paper, this lifted a coding agent to 91% on the HumanEval benchmark, against 80% for GPT-4 at the time.

"Many people have since reinvented it, but the basic technique is in a paper that a second-year college student wrote in 2023," Northeastern professor Arjun Guha told Business Insider.

Shinn then left college to join Sierra, the AI agent startup co-founded by Bret Taylor, as one of its first employees. There he kept publishing on agents, including τ-bench, a benchmark for how well agents use tools while talking with users.

What Went Wrong for Early Instinct Users?

An agent that acts for you can also act wrongly for you.

Reported fact: Some early Instinct users reported unexpected actions by the agent, the Wall Street Journal reported, according to VnExpress. Shinn said the company was taking the concerns seriously and changing the product to prevent unwanted actions.

One user made a more serious claim. On September 21, a user posted screenshots of the assistant talking about another person's financial document and middle name.

In them, the assistant apologized for "a serious glitch," RuntimeWire reported. Shinn denied that any user data was shared or that isolation was broken. RuntimeWire noted that the evidence doesn't show whether data crossed between accounts or the assistant simply made the details up. The claim remains unconfirmed.

Privacy came up earlier too. TechCrunch noted that Instinct's original privacy policy raised concerns before the company updated it.

These problems aren't unique to Instinct. Meta's Muse shared a user's home address with a stranger after one broad permission click, as we covered in our look at AI agent permissions.

What Safeguards Has Instinct AI Added?

Four, according to the company's funding announcement:

Safeguard	What it guards against
Isolated sandboxes	One task or user's environment reaching another's
Short-lived local credentials	Stolen or leaked logins staying useful for long
Identity-signed tool execution	Actions that can't be traced back to who or what started them
Hallucination detection before acting	The agent acting on details it made up

These are the right categories, and they match what we'd recommend for any agent with real access. The open question is how well they work. Instinct hasn't published test results, so treat them as promises until outside testing shows otherwise.

The last one is the most interesting for builders. Checking for made-up details before an agent acts, not after, is close to Reflexion's own idea: look at your work critically before you commit to it.

How Does Instinct Compare With Meta's Muse?

Same goal, very different starting positions.

Meta launched Muse in September as part of Mark Zuckerberg's push toward what he calls "personal superintelligence." Both products aim to move AI from conversation to action.

• Meta's advantages: huge computing resources, billions of users and built-in distribution through some of the world's largest social apps.
• Instinct's advantages: a head start with early adopters, a focused product and strong word of mouth.
• Instinct's limits: it's still invite-only, and it has faced heavy demand for computing capacity.

The rivalry has been public. Meta's chief AI officer, Alexandr Wang, has mocked Instinct on social media, once calling it "Insect," VnExpress reported. Shinn responded to one post with a smiley face and has called Muse a good product.

Should You Try Instinct AI?

If you get an invitation, it can save real time. Go in with care.

• Start small. Try low-stakes tasks first, such as research or a restaurant booking, before anything that spends real money.
• Grant only what's needed. Connect the accounts a task requires, not everything at once.
• Ask for confirmation. Where you can, set it to check with you before it pays, books or sends messages for you.
• Watch incoming messages. Emails and texts from strangers can contain hidden instructions. Be careful with tasks that read your inbox.
• Review and revoke. Check what the agent did each week, and remove access you no longer need.

These rules apply to any personal agent, not just Instinct.

What Can Agent Builders Learn From Instinct?

Five lessons, whether you're building for consumers or for a business.

• Use reflection loops. Let agents check their own results and write down what went wrong before retrying. It's cheap, and it often works.
• Build memory carefully. Remembering preferences makes an agent useful, but stored personal data raises the stakes of every mistake.
• Isolate every task. Run each user's work in its own sandbox, so one session can't see another's data.
• Keep credentials short-lived. Give agents logins that expire quickly and only cover what the task needs.
• Check before acting. Verify facts, amounts and recipients before the agent sends, buys or books anything.

Testing matters too. Our guide to AI agent evaluation covers how to build test sets from real tasks, and our agent ops playbook covers when an agent should hand work to a person.

What Does the $10 Billion Valuation Say About the Market?

That investors think personal agents could be the next big consumer platform.

Investors are betting on agentic AI: systems that do tasks, not just answer questions. An agent that books travel and buys things also sits on top of real spending. That's why transaction volume gets so much attention.

Our view: the valuation prices in a lot of future success. Instinct must scale computing capacity, prove its safeguards work and hold its lead against Meta's reach. None of that is guaranteed. But the demand is real, and it shows how much people want help with everyday tasks.

There's also a cost question. An agent that browses, makes calls and runs a computer for every user needs far more computing power than a chatbot. Instinct has already struggled to keep up with demand for capacity, VnExpress noted.

For teams building agents inside companies, the message is the same at a smaller scale. People will use agents that save them real time, as long as they can trust what the agent does with their accounts.

How Van Data Team Helps Teams Build Agents People Trust

We build AI agents that take real actions safely. That means reflection and checking loops, memory with clear limits, sandboxed tasks, short-lived credentials and approval steps for anything that spends money or shares personal data.

Instinct AI shows how fast a useful agent can grow, and how quickly trust becomes the hard part. If you're planning an agent that acts for your customers or staff, our AI governance guide is a good place to start.