The New Front Door

The Competitive Battleground Is Shifting from Technology to Trust

Every major technology company is racing to build a more capable personal agent. Most of the attention is on who can build the best one. But the more consequential race may be who can earn enough trust for people to give that agent a full picture of their lives and let it act on their behalf.

Trust is what gives an agent access to the context, preferences, and priorities that shape a person’s decisions. We call this the Human Context Layer—the structured, permissioned record that lets an agent act on a person’s behalf rather than merely answer their questions. It could become the substrate for much of the agentic internet, with platforms that earn durable, safely governed access to it becoming the new front door to people’s digital lives.

Major platforms are already building personal agents around this idea. On August 10, 2026, Meta laid out a vision for 24/7 personal agents serving its 3.6 billion daily users. With permission, these agents could draw on personal context across finances, careers, health, and relationships, with privacy protections similar to WhatsApp encryption. That vision is no longer hypothetical: Meta launched the consumer product, Muse, on September 8, 2026 – a task-executing agent that runs on a dedicated cloud VM and can book, shop, and negotiate on a user's behalf, free for most uses with $20/$100 monthly tiers for more. Google’s Gemini Spark and the open-source project OpenClaw currently operate on the same premise in narrower forms.

What remains unresolved is how and more importantly why, people grant agents access to their most sensitive context while retaining control over it, including the ability to take that context with them if they leave. And how that problem is solved may matter far more to the personal agent race than any incremental gains in model performance.

The Human Context Layer: What Makes an Agent Personal

An agent is only as useful as its understanding of the person it represents. The Human Context Layer draws on six domains of personal information, ultimately understanding and predicting context, needs, opportunities, and preferences:

• Communications (email, messages, relationship history)
• Temporal data (calendar, schedules, recurring patterns)
• Financial context (accounts, spending patterns, delegated spending authority)
• Health and wellness data (wearables, health records where permitted)
• Professional identity (work tools, roles, institutional affiliations)
• Behavior (browsing and purchase history, location, goals)

With this information, a person’s agent could notice a flight delay before they wake up, rebook a connecting train, and move a 9 a.m. meeting based on permissioned access to their calendar and travel loyalty accounts. But knowing where the person needs to be is only part of what informs the decision. The agent could also understand why this particular trip matters, how much they are willing to spend to stay on schedule, and which tradeoffs make sense in that situation.

The underlying architecture is not entirely new. The Solid project's personal data stores describe a similar architecture from a data-sovereignty standpoint, as does a broader body of research on personal information management systems. That architecture centers on personal data held in user-controlled stores that grant time-bound, revocable access to applications and agents. Context engineering and the agent memory layer built by companies such as Mem0, Zep, and Letta cover adjacent technical ground. What the Human Context Layer adds is a lens of useful application, treating that substrate as an asset that benefits people as much as the businesses serving them, rather than primarily a privacy right. That commercial race is already visible in this year's product launches: Meta's Muse, xAI's Grok Bot, Gemini Spark and Instinct from Spear Street Technology are all early attempts to monetize a version of this same context substrate, each wagering that owning a person's day-to-day requests creates end user value and product stickiness.

Google’s Personal Intelligence is one example of how personal context can inform an agent. Originally launched in January 2026, it can now connect Gmail, Photos, YouTube, and Search data to Gemini on an opt-in basis, with Calendar data added later. Answers draw on context Google already holds, and the assistant can handle light tasks across the Google ecosystem without being briefed first. It’s a narrow, single-ecosystem version of the Human Context Layer. But that context remains tied to the ecosystem, raising a broader question: who controls the Human Context Layer, and will it be portable across agents?

Portable Trust vs. Platform Lock-In

There are two very different ways this could play out:

• Context is portable. A person builds one rich profile and carries it between agents the way they carry a phone number between carriers. Switching platforms remains a real choice, without having to rebuild years of accumulated context elsewhere.
• Context is siloed. Each platform builds its own walled version of the Human Context Layer, making it painful enough to rebuild elsewhere that people rarely leave.

The two approaches differ on who controls a person’s context and who/what consumers trust. Tim Berners-Lee’s work on Solid argues for user-controlled data, while Mark Zuckerberg’s vision for Meta keeps that context within the platform with strong privacy protections. Both aim to make advanced capabilities broadly available, but they take different approaches to where that context lives. Consumers trust AI today in large part due to its perceived objectivity and may only choose to enable sharing context where it is truly beneficial to them.

For now, the second scenario seems more likely. Most existing architectures are designed to deepen trust within a single ecosystem, not make personal context easy to carry from one to another. That could change with regulation, a move by a challenger, or simply consumer resistance. While many consumers are already using AI-assisted shopping today, a January 2026 Clutch survey found that only 4% would let an agent complete a purchase on its own. That hesitancy is already playing out in real time. Instinct's rapid rise has been shadowed by scrutiny of its terms of service, which grant a perpetual, irrevocable license to user data including screen captures and keystrokes, and by an early incident in which the agent sent an email without the user's explicit permission – a live example of capability outrunning the trust infrastructure this paper argues is the real competitive battleground. If consumers are still hesitant to hand over the transaction itself, they may be even less ready to give any platform access to their personal context, much less to share it further.

Who Will Hold the Context?

At least eight platforms are building personal agents, each competing to become a trusted custodian of some version of the Human Context Layer. Google, Apple, Amazon, and Meta each bring installed bases above a billion users, though only Google and Amazon have something usable outside a beta today. Both Apple's Siri AI and Meta's agent, Muse, launched in September 2026. OpenAI, Perplexity, Anthropic, and xAI are competing more on architecture than distribution.

A wave of independent startups is entering from outside the incumbent platforms entirely. Instinct, a text- and call-based personal agent from Spear Street Technology, went from roughly a $500 million to a $2.5 billion valuation within weeks in August 2026, drawing intense investor interest for its ability to book, cancel, and negotiate on a user's behalf with minimal setup. xAI has also entered directly with Grok Bot, always-on agents bundled through its Cursor integration. Perplexity, for example, has integrated at the OS level with Samsung’s Galaxy S26, giving it access to an installed base it doesn’t own. Google has gone further at the OS level: Android 17 embeds “Gemini Intelligence” directly into the operating system, with an AppFunctions API letting third-party apps expose actions Gemini can execute on the user's behalf without switching apps — pushing the Human Context Layer down into the OS itself rather than a single app. Independent players such as Brave’s Leo browser, Fellou, and Hark are entering the market as well.

Protocols sit underneath these platforms and could determine whether agents interoperate or keep users tied to one ecosystem. Anthropic’s Model Context Protocol is one example, with monthly downloads growing from 97 million to nearly half a billion in five months. Google’s Agent Payments Protocol (AP2), an open protocol launched with more than 60 financial and technology partners, gives agents, merchants, and payment providers a shared way to transact across platforms.

How Personal Agents Change the Customer Journey

As agents gain access to more of a person’s context, they can take on more decisions on that person’s behalf. The agent increasingly sits between the person and the internet, changing how customers discover businesses, choose products, and complete everyday tasks:

Discoverability is being rebuilt around the agent, not the search box.

Zero-click searches already account for 68% of Google searches, rising to 83% when an AI Overview appears and above 90% in AI Mode on informational queries. A 5W Public Relations study found that the overlap between what ranks highly in traditional search and what AI systems cite fell from 70% to under 20% in a year. The businesses that appear in an agent’s answer will increasingly differ from those ranking on page one.

Commerce is shifting from browsing to intent fulfillment.

Google's Universal Commerce Protocol (UCP), backed by a coalition of retailers and payment providers, offers an early look at agent-driven commerce. Announced in January 2026, Google expanded the protocol two months later with multi-item carts, real-time catalog queries, and Identity Linking, which lets shoppers connect their loyalty accounts so an agent can factor in member pricing and benefits when making a purchase. And UCP is already moving beyond retail, with support for hotel bookings through companies such as Booking.com, Expedia, and Marriott, and food delivery through DoorDash and Uber Eats. Other companies are also adapting their commerce experiences for agents. Walmart’s ChatGPT app supports account linking, loyalty, and Walmart payments. Shopify’s Agentic Plan makes merchant products available across ChatGPT, Gemini, and other channels. And Amazon has brought Alexa for Shopping into its search experience.

OpenAI's Agentic Commerce Protocol (ACP), built with Stripe, plays a similar role for ChatGPT: it lets merchants such as Target, Sephora, Best Buy, and Walmart surface products for discovery, with checkout increasingly handled through the merchant's own ChatGPT app rather than inside the chat itself.

​​The important question for businesses is what actually influences an agent’s choice. When an agent chooses on a person's behalf, it weighs machine-readable attributes such as price, inventory, delivery commitments, return terms, feed freshness, and increasingly loyalty status. These details can directly affect the outcome a customer receives. Loyalty programs, long used primarily for retention, can also impact what an agent selects. A merchant that makes its loyalty benefits legible to an agent could win a comparison against a competitor with stronger brand equity but less accessible data.

Utility is where personal agents start to earn trust.

Before people trust an agent to spend their money, they may first let it manage a calendar, draft an email, or restock the household. Each of those tasks gives the agent more context about the person, adding to the Human Context Layer and making it more useful. That everyday utility is the pathway to build trust, making people more willing to grant permission for more consequential tasks, including commerce.

The Design of a Trusted Agent

Four architectural principles are emerging, with permissioning becoming part of the agent experience rather than a compliance checkbox:

• Human-in-the-loop approvals: Anthropic’s MCP OAuth requires approval above defined transaction thresholds, while OpenAI’s discovery-first pivot returns checkout to the merchant’s own site.
• Granular permissioning: Spending mandates, merchant-category controls, and limits a person can revoke at any time are being embedded at the network level by Mastercard’s Agent Pay and Visa’s Trusted Agent Protocol.
• On-device and locally run processing: Apple’s Private Cloud Compute keeps sensitive processing within a controlled device environment, while Meta’s Muse Glimmer is designed to run locally.
• Portability by design, not regulatory mandate: An exportable context layer would let people take their accumulated context with them when they switch platforms. Treating portability as a feature could put platforms ahead of the GDPR's data portability rules, the EU Data Act's device-data provisions, and whatever follows.

What Leaders Should Do Now

Much of the conversation around personal agents focuses on capability: which model reasons better, browses faster, or completes more tasks per dollar. But as those capabilities become more comparable, the model itself may matter less to businesses than what sits around it. For business leaders, three areas deserve attention:

• Structure for agent discovery, not just search ranking. Structured product data, open APIs, and relevant protocols increasingly determine whether a business appears in an agent’s answer. The businesses an agent recommends may differ from those ranking on page one, so companies should structure their products, services, and information around the needs and use cases and moments of demand they want to win.
• Make commercial terms machine-readable. Price competitiveness, loyalty status, inventory accuracy, and delivery commitments are increasingly part of how agents compare options. Benefits an agent cannot read may as well not exist.
• Build consent and context architecture deliberately. Permissioning granularity and revocability are becoming competitive attributes rather than compliance overhead, and they are difficult to retrofit once customer relationships are established. For the people on the other side of that architecture, this means making clear what an agent can do without asking, what it must check first, and what data it's using to decide – control that has to be visible, not just technically present.

A New Intermediary in the Customer Relationship

The Human Context Layer could change how businesses win customers. When an agent knows enough about a person to make decisions on their behalf, the customer is no longer the only one businesses need to reach. They also need to be understood by the agent making or influencing the choice. A product that fits the person’s needs and preferences may have an advantage over one with a stronger brand or search position if the agent can recognize that fit.

Businesses have spent decades investing in understanding their customers. Personal agents could change where that knowledge sits. An agent may have the context it needs to decide which business, product, or service fits best. If people trust the agent with that context, businesses will have to win the agent’s recommendation to win the customer.