When Agents Become the Customer: Part 2

Welcome back to the consumer AI agent train. One question we keep coming back to is whether consumers will eventually have one agent or many.

One Agent, or Fifty?

We don’t think the answer is either extreme. It seems unlikely that we want to individually manage fifty different AI agents: a travel agent, shopping agent, healthcare agent, scheduling agent, financial agent, restaurant agent and so on.

It also seems unlikely that one company will build the best possible agent for every domain. Some, like travel, require such depth and complexity that a dedicated, specialized agent may become necessary (think: Miso, Voyage Labs, and others). Our current guess is something in between: one agent that does a lot, but not everything, perhaps like a true chief of staff.

A generalist becomes your primary interface – the agent that understands who you are, knows your history, has broad permissions and handles the long tail of everyday tasks.

Behind it, might sit an ecosystem of specialists. Imagine the following task: “Plan my daughter’s birthday weekend.”

Your primary agent might know what she likes, what you’ve done before, which weekends you’re free, and approximate budget. But it may also delegate discovery to a dining agent, presents to a gifting agent, etc. As the consumer, you don’t necessarily need to know what’s happening behind the scenes. And over time, this starts to look less like today’s app ecosystem and more like a labor marketplace of software.

The Agent Marketplace May Have No Storefront

We are also skeptical that we will discover agents the way we discovered apps (though platforms like Agent Store make an interesting case). The App Store made sense because applications were destinations themselves. Agents are different. If I already have a trusted primary agent, why should I have to figure out which flight-search agent, tax agent or restaurant-booking agent it should use? That's work.

The future “agent marketplace” may therefore be largely invisible. Your primary agent could automatically discover, evaluate, and delegate to specialized agents based on the task at hand. And these agents may compete on price, reliability, permissions, reputation, latency or success rate.

But supply itself has to become agent-ready. If I tell my agent, “Find me a house cleaner for Thursday afternoon under $150,” it should not need to behave like a human – searching websites, filling out forms, calling businesses, waiting on hold. Instead, the cleaner, marketplace or service provider should expose its availability, price, policies, reputation and transaction rails in a way an agent can understand and act on directly.

So the marketplace has two layers:

• Delegation: my generalist agent finds the right specialist agent to help complete a task.

• Transaction: my agent interacts directly with the agent, system or infrastructure representing the supplier.

The UX is one sentence; the infrastructure behind it is an entire economy. There will be winners that develop the best consumer agents. There will be equally large opportunities in building the infrastructure that helps the rest of the economy become legible to those agents.

What Actually Becomes Defensible?

We mentioned before that the underlying intelligence may be one of the least defensible parts of a consumer agent. Models are improving quickly. Harnesses will converge. Features will get copied. Even traditional data lock-in may prove weaker than expected: migrating your preferences, purchase history, or context is exactly the type of tedious admin work that AI itself is good at.

The more durable assets may instead be:

• Distribution. Are you already present where users naturally express intent?

• Context. Do you understand enough about a person's life to act without repeatedly asking questions?

• Trust. Are users comfortable giving you the permissions required to be genuinely useful?

• Reliability. When someone asks you to do something, do you actually get it done?

• Network effects. Does the product become more useful as friends, family, merchants or other agents join?

The Race to the First Magic Moment

The capability curve isn't waiting for the adoption curve. Even if frontier models stopped improving tomorrow, there are probably years of product innovation left simply in getting today's capabilities into the hands of more people. That is the opportunity inside the diffusion gap. And consumer AI may not cross it through a radically new interface. It may cross it through the most boring interface imaginable.

A phone number, a text box, and a message: “Can you take care of this for me?”