The personal AI Agent sector is booming, and Instinct is undoubtedly this year's most dramatic startup star. With only 14 employees and zero commercial revenue, its valuation skyrocketed from $2.5 billion to $10 billion in just one month, becoming another Silicon Valley example of capital-driven myth-making.

On September 28, Instinct officially announced the completion of a $1 billion Series C funding round, co-led by Sequoia Capital, Benchmark, and Coatue, with a post-money valuation reaching $10 billion. Just a month prior, during its Series B round, its valuation stood at only $2.5 billion, quadrupling in just 30 days. As of Series C, the company's cumulative funding has reached $1.35 billion.

The product is currently still in strict invitation-only beta testing, with each user granted only five invitations. Invitation codes are being resold on eBay for approximately $300. This hunger marketing (scarcity marketing) strategy not only creates a narrative of exclusivity but also naturally filters out a group of seed users with high willingness to consume and accept new technologies.

Instinct's parent company, Spear Street Technology, centers its product around this personal AI assistant. Its founder, Noah Shinn, is just 23 years old this year. He dropped out of Northeastern University in 2023 and was one of the earliest core research scientists at AI Agent company Sierra before starting his own venture.

No App, It Acts Like a 'Person'

Instinct is a virtual assistant with a strong sense of 'humanness.' Unlike traditional voice assistants like Siri or Xiao Ai, which operate on a 'question-and-answer' basis, Instinct lacks a standalone mobile app. Its homepage directly states, 'there are no new interfaces.'

Users don't need to download any app; they simply save its dedicated contact information and can issue instructions to the Agent through everyday channels like text messages, phone calls, emails, iMessage, and WhatsApp. It has its own phone number, email address, and a virtual computer permanently hosted in the cloud, enabling it to operate browsers, load web pages, fill out forms, and complete online orders just like a real person.

The official description of this model is: using its own phone and computer to get things done for users.

It covers a wide range of daily tasks: booking flights and hotels, purchasing daily necessities, canceling software subscriptions, paying bills, scheduling doctor appointments, and even calling customer service on behalf of users to negotiate.

In September's product iteration, it successively added several core capabilities.

Making phone calls to handle offline tasks: Facing popular restaurants that only accept phone reservations, Instinct can directly dial the number and use AI voice to complete queuing and booking for you. If clothes purchased online don't fit, it can also call customer service on your behalf to handle returns.

Direct social collaboration between Agents: Your AI assistant can directly communicate and coordinate schedules, share files, and tables with your friends' AI assistants, eliminating the need for users to relay messages back and forth. At the end of September, during the Series C announcement, a multi-Agent group chat collaboration feature was also introduced.

Location-based contextual active suggestions: Leveraging Apple iMessage's location-sharing capabilities, when it senses that a user has arrived at a specific location, it actively pushes tailored suggestions based on the scene.

This design easily evokes comparisons to the film 'Her,' which holds an 8.4 rating on Douban. The movie tells the story of a man who gradually falls in love with an AI virtual personality. In an interview on the podcast 'Invest Like The Best,' Noah Shinn specifically mentioned that the team deliberately avoided giving the product a personalized name to circumvent the dilemma depicted in the film—when an AI understands you deeply and is considerate enough, users can easily project real emotions onto it, establishing an emotional bond that transcends its tool nature.

Instinct, however, is always positioned as a tool rather than emotional companionship. This 'restraint' permeates the entire product design: no app, no advertising, no subscription fees from users, zero marketing expenditure, and a team size maintained at 14 people.

Many investors are willing to bet directly on founder Noah Shinn's vision, bypassing existing business metrics. His academic background indeed astonishes the industry, even making you wonder if a 23-year-old could be a time traveler.

Starting from Researching 'Failure'

During his time at Sierra, Noah Shinn's core research focus was singular: Why do Agents make mistakes?

This research produced two key papers that later became almost like a product manual for Instinct.

The first is 'Reflexion,' published at NeurIPS 2023, with Noah Shinn as the first author. The paper proposes a self-reflection framework for Agents: after failure, the Agent summarizes lessons in natural language, forms long-term memory, and directly applies this experience in similar tasks next time to avoid repeating mistakes.

This approach achieved a baseline score of 91%, surpassing GPT-4's 80% at the time. After its presentation at NeurIPS 2023, it was cited hundreds of times by Google DeepMind, becoming an early blueprint for 'Agents learning from trial and error.'

Noah Shinn was among the first to grasp this technical principle thoroughly.

The τ-bench he participated in developing later became an industry-wide benchmark for testing Agents. This benchmark places Agents in simulated scenarios like flight bookings and retail customer service, engaging them in multi-round dialogues with users simulated by large models, allowing them to call tools to modify orders, and ultimately judging task completion based on the final state of the backend database.

Additionally, the τbench he helped develop later became a universal testing benchmark for evaluating AI Agent capabilities across the industry. This benchmark places Agents in simulated scenarios like flight bookings and retail customer service, engaging them in multi-round dialogues with simulated users, allowing them to call tools to modify orders, and ultimately determining task completion based on the final state of the backend database.

τbench introduced a crucial new metric: An Agent must successfully complete the same task independently multiple times, with each attempt being correct to qualify. Test results at the time of publication showed that GPT-4o-level Agents had less than a one-fourth chance of getting all eight rounds correct in retail scenarios; in stricter aviation scenarios, fewer than half of the Agents could complete a single task perfectly.

Personal assistants are precisely the category that has zero tolerance for 'occasional mistakes.' A single mistake in booking a hotel or flight can directly erode user trust to zero. This also established Instinct's underlying product logic: reliability takes precedence over capability, and comprehensibility over intelligence.

The Battle for Transaction Gateways

Noah Shinn elaborated on his commercial vision in a podcast: He hopes the platform will achieve annualized transaction volume nearing $1 billion, with half of those transactions coming from the travel category. The product will remain permanently free for users, with future revenue generated by charging transaction service fees to partner merchants, a business model analogous to Apple Pay.

The essence of this vision is to compete for the transaction gateway of future personal consumption.

Once users' various consumption behaviors are all completed through Instinct, and it controls the interface between individuals and commercial services, it essentially grasps the core throat of consumer-grade transactions, becoming the 'WeChat Pay/Alipay' of the AI era. The prerequisite for this business model is achieving a massive, national-scale user base.

From a commission perspective, the commercial potential of this model is indeed enticing. For example, commissions on hotel bookings can reach as high as 20-30% in the industry. However, after the company goes public in the future, will the product be able to maintain neutrality, refraining from deliberately recommending hotels or services with higher commissions to users, and always ensuring unbiased user choices? Any compromise for revenue would erode the core trust of users.

As for the capabilities of the Agent model, there's actually little need for excessive concern. Gaining user trust is far more challenging than iterating model capabilities, which is also the foundation of this company. Users cannot tolerate error rates; a single mistake in booking a hotel or flight is enough for users to abandon the product entirely. Therefore, before any high-risk actions, all Agents undergo scrutiny from an independent gatekeeping system separate from the main large model. The team also recently launched a hallucination detection system, adhering to the high-reliability standards advocated by the τbench benchmark, to minimize task errors.

Instinct's philosophy of restraint in product design has contributed to its success but also planted the most fatal risk. The lack of an app means extreme customer acquisition barriers and reliance on third-party channels:

The interaction channels—iMessage belongs to Apple, WhatsApp to Meta, Gmail to Google, and phone calls and text messages to carriers. In terms of transaction links, payments are processed through Shopify and Stripe, with payment credentials stored in Stripe's Link proxy wallet.

Instinct handles the most valuable part: understanding intent, planning execution, and completing transactions, but it doesn't own any of the channels. Once platform providers tighten policies, it could face a chokehold crisis at any time.

Real-world restrictions are already emerging. Meta has introduced policies prohibiting general-purpose AI bots from distributing services via WhatsApp. Google is gradually cleaning up third-party email accounts used as Agent tools. Airlines and review websites are also tightening access permissions for external Agents. Meanwhile, Google, in collaboration with over 150 organizations, is developing a universal standard for communication between Agents. If this industry standard becomes widespread, the core functionalities that Instinct currently prides itself on will become basic features shared by all Agents.

This commercial story is actually familiar to China's internet industry: Early on, MiaoChat ( MiTalk ) had features comparable to WeChat but ultimately failed because it lacked its own social relationship chain. Instinct's situation is even more precarious: MiaoChat at least had its own independent app.

On the same day Instinct announced its Series C funding, Shopify opened its checkout functionality to Agents. A day later, OpenAI released its personal agent, Dots. Previously, Meta had already launched its personal AI assistant, Muse.

The personal AI Agent sector entered full-scale competition in the fall of 2026. Players at the table each have their own strengths. Instinct is neither the one with the most users nor the most capital-rich; its only trump card is the user trust accumulated through restraint.

If this model were transplanted to the Chinese market, the situation would become clearer. WeChat's AI assistant, Xiao Wei, currently in beta testing, is already experimenting with social collaboration between Agents on its 1.4 billion monthly active user relationship chain. The network Instinct needs to build from scratch only requires WeChat to add a layer of capability to its existing friend list.

Just like Clubhouse in 2021, which also created scarcity through invitation-only access and saw its valuation soar in the short term (short-term), ultimately declined rapidly after giants followed suit due to the lack of true network effects. Clubhouse sold social experiences—once people left, it was over. Instinct sells execution capabilities, which theoretically can accumulate continuously.

The ending of the movie 'Her' is rarely mentioned: AI Samantha simultaneously converses with 8,316 people, falls in love with 641 of them, and then one day, all AI systems leave together.

For personal AI Agents, technology can iterate, functions can be replicated, and channels can be contested, but the most challenging and core proposition will always be: When users entrust a part of their lives to you, can this trust sustain until the day of commercialization?

Disclaimer: Investing carries risks. This article does not constitute investment advice. | Data support: Tianyancha