The Personal Agent Race Is Here

What Instinct reveals about the next interface to the internet

🎧 Listen on NativePod: Noah Shinn — Building Instinct: The Personal Agent, from Invest Like the Best with Patrick O'Shaughnessy.

For years, AI assistants have promised to make life easier. Most still wait for a prompt, produce an answer, and hand the remaining work back to the user. Instinct founder Noah Shinn is pursuing a more consequential idea: an agent with its own phone, email address, and computer that can act across the internet on a person's behalf.

In this episode of Invest Like the Best, Shinn and Patrick O'Shaughnessy explore what happens when AI stops being a destination and becomes an operating layer for everyday life. The discussion ranges from travel and commerce to trust, safety, business-model alignment, and the extraordinary compute requirements of an agent that works even while its user is asleep.

The Product Thesis: No New Interface to Learn

Instinct's most important design choice may be what it removes. A user can text, call, or email the agent through channels they already understand. The system can call back when something is urgent, work through websites on its computer, and follow a task from intent to completion.

That makes the experience feel less like opening software and more like delegating to a capable operator. A request such as “I need to be in New York tonight” can imply flights, seat preferences, hotel selection, ground transportation, calendar updates, and payment. The interface is one sentence; the workflow behind it spans many services.

This distinction matters. The next consumer AI breakthrough may not come from exposing more model capabilities. It may come from hiding complexity while preserving control.

From Answers to Outcomes

The episode's strongest examples show why action changes the value of AI:

• A digital wardrobe can catalog clothing, understand a user's proportions, propose complete outfits, preview new purchases, and place an order.
• A financial cleanup task can identify unused subscriptions, log into each service, complete cancellation flows, and report the savings.
• A travel request can combine location, loyalty preferences, seating, hotels, ride-hailing, and calendar constraints into one coordinated plan.
• A small business can give the agent an ongoing inventory target instead of repeatedly issuing purchase instructions.

The common thread is not automation for its own sake. It is the removal of logistical work between an intention and a completed result.

A Trusted Network of Agents

One of Instinct's more original ideas is a network in which two people's agents can coordinate directly. Scheduling is the obvious example: both sides already know their users' calendars, preferences, and constraints, so they can find a time without a long exchange of messages.

But the deeper concept is a weighted trust graph. A spouse might receive broad access, a colleague might see only work-calendar availability, and another contact might receive a narrowly scoped permission. The connection is not simply “friend” or “not friend”; it encodes what each relationship is allowed to know and do.

That creates a new social layer for software. An attempted boundary violation is not only a technical event. It can be surfaced to the user as a breach of interpersonal trust. If personal agents become common, permission design may become as important as the intelligence of the underlying model.

Trust Is the Adoption Funnel

Shinn describes trust as something users grant gradually. According to figures he shares in the episode, roughly 40% of users had provided a personal credit card after about three weeks, and retention was around 80% after a user connected at least one sensitive source.

These are company-reported metrics, not independently verified benchmarks, but they reveal a useful product lesson: for an action-taking agent, the activation funnel is really a trust funnel. The more useful context a person shares, the more proactive the agent can become. The more reliably it acts, the more trust it can earn.

The loop is powerful—and dangerous if mishandled. Users need to understand what the agent can access, grant permissions at their own pace, and revoke them at any time.

Business Models Are an Alignment Problem

An agent that can persuade and purchase on a user's behalf cannot treat monetization as an afterthought. If advertising determines recommendations, the agent may optimize for a merchant rather than the person it represents.

Instinct's proposed model is closer to payments and distribution infrastructure: keep the consumer experience free or affordable, then charge merchants for incremental transactions. Shinn says the platform was approaching $1 billion in annualized transaction volume at the time of the conversation, with travel representing about half.

The strategic question is whether a personal agent can create enough demand and reduce enough friction to earn a share of a transaction without compromising its duty to the user. That is not merely a pricing decision. It is the constitutional design of the product.

From the Attention Economy to the Intent Economy

Personal agents could rearrange the economics of the internet. Services that make money from fulfilling genuine demand—transportation, hotels, restaurants, commerce—may benefit as purchasing friction approaches zero. Products that depend on keeping people scrolling or making simple tasks tedious face a different future.

The episode offers a useful way to analyze any digital business: how much revenue comes from delivering the underlying service, and how much comes from controlling user attention inside an interface?

An agent can still send a customer to Uber, an airline, or a hotel. But it may absorb the browsing, comparison, form-filling, and upselling surfaces that those companies currently control. The likely outcome is not universal displacement. It is a renegotiation of who owns demand, distribution, and the customer relationship.

Product Craft Beyond Model Capability

Shinn repeatedly emphasizes comprehensibility over raw capability. Users should be able to predict what will happen when they ask for something. Messages should minimize cognitive load, put the most important information first, and adapt to how each person prefers to communicate.

That leads to a different development process. Soft qualities such as trust, restraint, timing, and social awareness cannot be measured by a single task-completion score. Instinct reportedly stages product changes from the founder, to the team, to a small early-access group, and only then to broader release.

The lesson for AI product teams is clear: intelligence may become abundant, while judgment and taste remain scarce.

The Compute Problem No One Sees

A proactive agent consumes compute differently from a chatbot or coding tool. It may wake before the user, inspect a calendar, monitor traffic, prepare for a meeting, notice a deadline, or complete background research—without receiving a new prompt each time.

Shinn says Instinct was growing around 10% per day through invitations and word of mouth, with no marketing spend. At that rate, capacity planning becomes unusually hard because compute often needs to be reserved months in advance. Buying too little breaks the product; buying too much can multiply costs.

The team attempts to lower cost by matching infrastructure to the workload. Not every task needs frontier-model latency. Batch processing and specialized deployments can make background work several times more efficient. Many small optimizations compound into the economics required for a broadly accessible agent.

Safety Must Be Architected, Not Promised

Giving software access to email, payments, calendars, and location expands the consequences of every error. Instinct's described approach separates the acting agent from systems that inspect incoming content and review intended actions before execution.

That architecture resembles defense in depth:

• Screen external content before it reaches the agent.
• Detect prompt injection and malicious instructions.
• Verify sensitive actions before tools execute them.
• Use an independent monitor whose incentives differ from the acting system.
• Keep permissions visible and revocable by the user.

No system can make risk disappear, but separating execution from oversight is a stronger foundation than relying on the same model to police itself.

The Interface After Apps

The long-term vision is more radical than a better messaging bot. Voice could reduce interaction to a hardware button or an always-available earbud. When information needs a richer surface, the agent could generate a temporary interface—a travel plan, wedding dashboard, or research workspace—rather than asking the user to install another permanent app.

Eventually, the unit of interaction may shift from tasks to goals: improve a mile time over four months, keep inventory within a target range, or help a family coordinate a complicated week. The agent would choose and sequence actions while the person retains authority over the outcome and boundaries.

What to Watch Next

Instinct is still an early, invite-only product, and the company metrics in this conversation are claims made by its founder. The bigger ideas are nevertheless worth tracking:

1. Can personal agents earn durable trust without forcing users to surrender control?
2. Will merchant-funded economics remain aligned with user intent?
3. Can security systems keep pace with expanding permissions and attack surfaces?
4. Do incumbent platforms cooperate with agents that reduce time spent in their apps?
5. Can the industry supply enough compute at a price that makes personal agents broadly available?

The personal-agent race will not be won by the model with the longest feature list. It will be won by the product people trust to act when they are not watching.

Hear the full conversation in English or Chinese on NativePod.