For more than a year, Mark Zuckerberg has been describing a future built around personal superintelligence: AI that knows an individual’s context, understands what matters to them, and works continuously on their behalf.

On September 8, Meta finally put a consumer product behind that ambition.

Muse is a personal AI agent that can send emails, book travel, fill out forms, make purchases, negotiate on a user’s behalf, and keep working after the user closes the app. It can connect to services such as Gmail, OpenTable, Ticketmaster, Shopify, Spotify, and Stripe. Meta says it will eventually extend Muse into its AI glasses, giving the agent an even more persistent place in daily life.

Muse did not arrive in a vacuum. A new category has been taking shape quickly.

Town has built a persistent assistant that learns how a person works across email, calendar, documents, Slack, and other tools, then takes action rather than waiting for the next prompt.

Instinct has generated extraordinary buzz by making the experience even more personal: users can text or call it and ask it to book travel, arrange appointments, buy things, follow up on forgotten tasks, and handle the administrative friction of everyday life.

Grok Bot takes a similar pattern into persistent AI “teammates” that have their own computers, can sign into applications, work across tools, and return when they need approval. Google has already begun rolling out Gemini Spark, its 24/7 personal agent inside Gemini, as part of what Google calls Personal Intelligence.

I have been using Grok Bot, Town and Instinct myself. That matters because the value proposition becomes much easier to understand once you experience it personally. You stop thinking about “using AI” and start thinking about handing something off.

You message the agent the way you would message a capable assistant. You give it an outcome. It works across applications. It remembers context. It comes back when it needs a decision. When it works well, the convenience is not incremental. It changes the interface between you and the digital world.

That is why Muse matters.

Meta is attempting to take an experience pioneered by smaller companies and put it in front of a mass consumer audience through an ecosystem that includes WhatsApp, Instagram, Facebook, and eventually AI glasses. It is one of the clearest attempts yet to make delegated intelligence a mainstream behavior.

And Meta is doing this at exactly the moment when trust is hardest for the company to ask for.

Less than two weeks before the Muse launch, Meta agreed to pay up to $18 billion to settle claims brought by 29 U.S. states over alleged harms to children on its social platforms. Meta did not admit liability. The settlement follows years of privacy controversies, regulatory scrutiny, litigation, and public debate over how the company collects, uses, and protects personal information.

Now Meta is asking consumers for something far more consequential than permission to personalize a feed.

It is asking for permission to act.

That tension sits at the heart of the consumer agent era. The more useful an agent becomes, the more context and authority it needs. The more context and authority it receives, the greater the consequences when something goes wrong.

The Age of Delegation

The first wave of generative AI was largely about access to intelligence.

Ask a question. Draft an email. Summarize a document. Generate an image. Analyze a spreadsheet. Write some code.

The human remained the last mile.

AI could draft the email, but a person sent it. It could recommend a flight, but a person booked it. It could analyze a contract, but a person signed it. It could suggest what to buy, but a person entered the payment information and completed the transaction.

Consumer agents remove more of that last mile.

Consider the difference between two requests:

“What flight should I take to New York?”

“Get me to New York Thursday morning. Use my preferred airline if the fare is reasonable. Book an aisle seat. Make sure I am home Saturday for dinner.”

The first request asks for advice.

The second delegates an outcome.

To fulfill it, an agent must interpret intent, understand preferences, check constraints, compare alternatives, make tradeoffs, interact with external systems, and potentially commit money.

That is a fundamentally different relationship with software.

The shift from AI that advises to AI that acts is a shift from intelligence to authority.

A New Interface for Computing

For most of computing history, people have adapted themselves to software.

We learned operating systems, browsers, menus, forms, applications, and workflows. We learned which site to visit, which app to open, which fields to complete, and which buttons to press.

Personal agents invert that relationship.

The user expresses intent. The agent determines how the digital world should be navigated to fulfill it.

Instead of opening an airline app, a hotel app, a calendar, a restaurant service, maps, and email to organize a trip, you increasingly tell an agent what you want. The agent interacts with the applications.

This is why the chat interface matters.

It is not simply another place to talk to AI. It may become a universal control surface for software.

Town, Instinct, Grok Bot, Gemini Spark, and now Muse are all approaching the problem differently, but they point toward the same behavioral shift: fewer application-level instructions, more outcome-level delegation.

The interface becomes intent.

From Personal Convenience to Personal Power

It is easy to describe these products as assistants. That understates what is happening.

A useful personal agent combines several capabilities that were previously separate: reasoning, memory, application access, credentials, payments, communication, and persistent background execution.

Put those together and the user gains a kind of digital leverage that previously required staff.

A good personal agent can remember what you forget, monitor what you cannot continuously watch, coordinate across systems, and keep working when you are doing something else.

That is an extraordinary consumer proposition.

It also creates an extraordinary concentration of access.

To become genuinely useful, the agent may need some combination of your email, calendar, contacts, files, purchases, travel history, location, subscriptions, financial information, health information, passwords, and communication patterns.

This is why the consumer agent race will not be won on model intelligence alone.

It will be won on trust.

The Trust Paradox

Muse makes the trust problem unusually visible.

Meta says each Muse operates inside a dedicated Secure VM. A separate Sentinel agent controls what Muse can send to the internet and can require user approval. Credentials are isolated so Muse can use them without seeing the underlying passwords or payment details. Users decide which services Muse can access and how much access it receives. Sensitive actions such as sending email or making purchases can require approval. Meta says users can inspect an audit trail of what Muse has done and what it plans to do.

Those are important protections.

They are also evidence of how different the agent problem is from the chatbot problem.

A chatbot getting something wrong may produce a bad answer.

An agent getting something wrong can produce a bad outcome.

It can send the email.

Spend the money.

Cancel the reservation.

Change the account.

Share the information.

Commit the company.

The governance question therefore changes from “Can we trust the answer?” to “Can we trust the action?”

That is a much harder question.

Shadow AI Is About to Become Shadow Agency

This is where the consumer story moves directly into the enterprise.

The first enterprise challenge created by generative AI was Shadow AI: employees using unsanctioned models and applications outside the organization’s approved technology environment.

Personal agents raise the stakes.

An employee can now have an external agent that remembers context, holds persistent credentials, accesses email and calendars, reads files, interacts with websites, and takes actions on the employee’s behalf.

That is not simply Shadow AI.

It is shadow agency.

Imagine an executive connecting a personal agent to a corporate inbox because it is excellent at triage. A salesperson allows one to manage follow-ups. An employee gives an agent access to a calendar, cloud drive, expense system, or procurement portal. A developer lets an agent operate inside production tools. The agent may be extraordinarily useful, but the enterprise may have no reliable record of what authority it has been given, what information it can see, where that information is processed, or what actions it can take.

This creates a category of risk that most existing AI policies were not written to address.

The question is no longer merely which model an employee is using.

The questions become:

What has the agent been authorized to access?

What can it do?

What credentials does it possess?

What can it communicate externally?

What can it purchase or commit?

What information is it retaining?

Which third parties can it interact with?

How is its authority revoked?

Can the company reconstruct what it did?

That is why consumer agents should already be on the boardroom agenda.

The Consumer Agent Becomes the Enterprise Agent

The same interaction pattern that makes personal agents compelling will make enterprise agents compelling.

Instead of asking AI to draft a sales email, an organization can authorize an agent to identify prospects, research them, decide which warrant outreach, personalize communications, schedule meetings, and update the CRM.

Instead of asking AI to analyze an invoice, an agent can reconcile it, approve it within a defined threshold, and trigger payment.

Instead of asking AI to recommend how to resolve a customer complaint, an agent can issue a credit.

Instead of asking AI to identify a supplier problem, an agent can negotiate with the supplier.

Instead of asking AI to flag an operational anomaly, an agent can change the configuration that caused it.

This is where the board’s governance frame must expand.

The important distinction is no longer simply whether an organization uses AI.

It is whether AI has authority.

Context Becomes Strategic

A chatbot can be useful without knowing much about you.

An effective agent cannot.

It needs context: preferences, relationships, obligations, history, constraints, priorities, and the exceptions that shape judgment.

For an individual, that might mean knowing which airline you prefer, when you will tolerate an early flight, which loyalty programs matter, who is traveling with you, and when convenience is worth paying for.

For an enterprise, the contextual layer is much more complicated.

Who owns the customer relationship? Which contracts contain unusual provisions? Which customers receive pricing exceptions? Which approvals are required? Which regulations apply? What happened the last time the organization faced a similar situation? Which executive has authority to make the call?

This makes context an increasingly important strategic asset.

Model capability will continue to matter. But as capable models become widely available, competitive advantage may shift toward the systems that possess the best context, the clearest authority, and the strongest evidence.

The flywheel is straightforward: better context can produce better decisions; better decisions can justify greater trust; greater trust can support greater delegation; greater delegation produces more context.

For boards, that raises a strategic question that sits alongside cybersecurity and model risk: who owns the contextual layer through which agents understand the enterprise?

When an Agent Can Spend Money

Muse crosses another important boundary: it can transact.

Meta has integrated Muse with Stripe’s Link wallet, allowing the agent to make purchases with user approval while protecting underlying payment credentials.

The significance goes well beyond easier checkout.

An AI system that can transact becomes an economic representative.

It can compare prices, negotiate, purchase, return, subscribe, cancel, rebook, and allocate resources. Eventually, agents will routinely transact with other agents.

A consumer travel agent may interact with an airline agent.

A procurement agent may negotiate with supplier agents.

A treasury agent may move cash among approved financial products.

An insurance agent may continuously evaluate coverage and pricing.

A sales agent may negotiate commercial terms within defined limits.

Once machines can commit resources, identity and payment are not enough.

The system must also establish scope of authority.

Who authorized the agent? What could it spend? What could it commit to? Which counterparties could it interact with? What required approval? When did its authority expire? Could that authority be revoked? What evidence proves that the transaction occurred within mandate?

These become operating requirements.

The Oldest Governance Problem Gets a Literal Agent

There is an important irony in the terminology.

Corporate governance has spent decades grappling with the principal-agent problem: how a principal ensures that someone acting on its behalf exercises delegated authority in accordance with the principal’s interests.

Shareholders delegate authority to boards. Boards delegate authority to management. Executives delegate authority throughout the organization. Policies, controls, incentives, reporting, audit, and accountability exist in part to ensure that delegated authority is exercised properly.

Agentic AI introduces a new actor into that chain.

This time, the agent is literally software.

That does not make traditional governance principles obsolete. It makes them more important.

Boards will need to know where authority originates, how it is delegated, how far it extends, how it is monitored, and how the organization can prove that it was exercised appropriately.

The technology is new.

The governance problem is not.

Human Authority Over the Loop

A common response to autonomous AI is to insist on a “human in the loop.”

That will be necessary for some decisions, but it cannot be the complete operating model.

If a human must approve every action, much of the value of autonomy disappears. If agents receive unrestricted authority, the risks become unacceptable.

The more useful model is bounded autonomy.

Humans establish objectives, permissions, limits, escalation thresholds, and accountability. Agents operate within those boundaries and return to humans when a decision exceeds them.

An organization might authorize an agent to issue refunds below $500, purchase approved products within a predefined budget, negotiate a contract but not execute it, move money among internal accounts but not externally, access a data set but not disclose it, or communicate with customers until a defined risk threshold is crossed.

The goal is not a human in every loop.

It is human authority over the loop.

That distinction becomes essential at machine scale. A person may make dozens of consequential decisions during a workday. An autonomous system can make thousands of decisions or take thousands of actions between management reviews.

Governance designed around episodic human approvals will not be sufficient.

The Delegation-of-Authority Matrix Becomes Software

Most large organizations already have a delegation-of-authority framework.

It defines who may approve expenditures, sign contracts, hire employees, make commitments, access information, or take other consequential actions.

Historically, those rules were designed for people and documented in policies, process maps, access-control systems, and approval workflows.

In an agentic enterprise, those rules increasingly need to become machine-readable and enforceable.

An agent should not simply be told to “help procurement.”

Its authority should be explicit.

Which systems can it access? Which data can it use? Which vendors can it contact? What can it purchase? Within what budget? Under which contractual terms? What requires escalation? How long does the authorization last? Who can revoke it?

This is where AI governance moves beyond principles and policy documents.

The organization needs controls that can travel with the agent and govern its behavior as it moves across systems.

Delegation itself becomes programmable.

Every Agent Will Need a Record

As autonomy increases, evidence becomes essential.

It is not enough to know what an agent was designed to do. Boards, management teams, auditors, insurers, customers, and regulators may need to know what it actually did.

Organizations should be able to reconstruct an agent’s identity, owner, operator, capabilities, permissions, models, connected systems, data access, approvals, decisions, actions, exceptions, incidents, and changes in authority over time.

After an incident, the organization should be able to answer basic questions with evidence:

What happened?

Which agent acted?

Under whose authority?

What did it know at the time?

Which controls applied?

Did it remain within its mandate?

Who approved the exception?

What was the outcome?

Auditability cannot be bolted on after autonomous systems become material to the business. It has to be part of the architecture.

This Is Also a Growth Question

The rise of agents is not only a risk and control story.

It is a strategy and growth story.

If consumers increasingly delegate decisions about where to stay, what to buy, which restaurant to visit, what insurance to select, which software to use, or which vendor to hire, companies will increasingly compete for the attention and preference of agents as well as people.

For two decades, companies optimized information for search engines. Then they optimized products and content for social feeds and mobile platforms.

Now they must consider whether an agent can accurately understand their products, prices, policies, inventory, terms, reputation, availability, and service levels.

The customer journey may increasingly begin with a machine acting on behalf of a person or organization.

That means every company will have to consider two customers: the human and the human’s agent.

For boards, the strategic question is therefore broader than “How are we using agents?”

It is also: “How will agents use us?”

Questions for the Board and Management

1. Where are AI systems already taking actions on behalf of the company, rather than simply advising employees?

2. Are employees using personal agents such as Town, Instinct, Grok Bot, Muse, Gemini Spark, or other tools with access to corporate email, calendars, files, credentials, or systems, and would we know if they were?

3. What authority have we delegated to AI systems, and is that authority explicit, bounded, time-limited, and revocable?

4. Which decisions should remain human, which can become autonomous, and who has the authority to draw that boundary?

5. Can we reconstruct and independently verify what an agent did, what information it used, why it acted, and what authority it possessed at the time?

6. Who is accountable when an agent acts outside its mandate, makes an unauthorized commitment, discloses sensitive information, or causes economic or reputational harm?

7. How will our company interact with customers, suppliers, employees, and counterparties that are increasingly represented by their own agents?

8. Are our governance and control systems designed for human-speed oversight, or for an environment in which machines may take thousands of actions between board meetings?

The Age of Delegation

Muse may succeed spectacularly or Meta may discover that consumers are not yet ready to entrust the company with this much of their digital lives.

Either outcome is less important than the behavior Muse is trying to normalize.

Give software context.

Give it credentials.

Give it tools.

Give it a goal.

Give it authority within defined boundaries.

Then let it work.

That is a very different model of computing from the one most organizations were built to govern.

For decades, corporate governance has been concerned with the delegation of authority from shareholders to boards, from boards to management, and from management throughout the enterprise.

We are now adding another participant to that chain: software.

The defining question of the agentic era will not be whether machines become intelligent enough to act on our behalf. Increasingly, they are.

The harder question is whether individuals and institutions can become disciplined enough about the authority they give them.