Every transformative platform eventually gets a marketplace. The iPhone got the App Store. The web got hosting providers and plugin ecosystems. Salesforce got the AppExchange. Shopify got its theme and app stores.
AI agents are about to get theirs. And the companies that build the winning marketplaces will define the next decade of software distribution.
## Why Marketplaces Are Inevitable
Right now, if you want an AI agent that does something specific, you have three options. Build it yourself, hire someone to build it, or hope one of the major providers offers something close enough.
That's exactly where mobile apps were in 2007. You could build your own software, hire a developer, or use whatever came preloaded on your phone. The App Store didn't just make distribution easier. It created an entirely new economy. Independent developers could build once and sell to millions. Consumers could discover solutions to problems they didn't know had solutions.
Agent marketplaces will follow the same trajectory. But with some critical differences that make them far more interesting.
## The Architecture Problem
Apps are self-contained packages. You download them, they run on your device, they store data locally or in their own cloud. The interaction model is simple: user opens app, app does thing, user closes app.
Agents are fundamentally different. They need to interact with your systems. Your CRM, your email, your databases, your APIs. A customer support agent needs access to your ticketing system, your knowledge base, your customer history. A code review agent needs access to your repositories, your CI/CD pipeline, your style guides.
This means an agent marketplace can't just distribute code. It has to distribute capabilities that integrate with your existing infrastructure. The marketplace needs to handle authentication, permissions, data access, tool configuration, and runtime orchestration. It is worth reading about [agent-to-agent commerce](/blog/agent-to-agent-commerce) alongside this.
Think of it less like downloading an app and more like hiring a contractor who comes with their own tools but needs access to your building.
## The Trust Layer
Here's where it gets genuinely difficult. When you install an app, you're giving it access to defined device capabilities. Camera, location, contacts. The permission model is well understood.
When you deploy an agent, you're giving it access to business-critical systems with the ability to take actions autonomously. A procurement agent with access to your purchasing system can spend real money. A customer communication agent with access to your email can damage real relationships.
The trust layer in an agent marketplace needs to be fundamentally more robust than anything we've built before. I expect it to include several components.
**Capability declarations.** Every agent explicitly declares what tools it needs, what data it accesses, and what actions it can take. No hidden capabilities. No scope creep.
**Sandboxed execution.** Agents run in controlled environments where their actions can be monitored, rate-limited, and reversed. You don't give the agent direct database access. You give it a controlled API with guardrails.
**Audit trails.** Every action an agent takes is logged immutably. Not just for compliance. For debugging, for trust building, for understanding what your agents are actually doing.
**Reputation systems.** Like seller ratings on Amazon, but for agents. How often does this agent complete tasks successfully? How often does it escalate to humans? What's its error rate? What do other organizations report about its performance?
**Certification tiers.** Different levels of verification for different risk levels. A newsletter-writing agent needs less scrutiny than a financial transactions agent. The marketplace should reflect that.
## The Economics
This is where I think most predictions go wrong. People assume agent marketplaces will mirror app store economics. 70/30 revenue splits, subscription models, one-time purchases. That's too simple.
Agent marketplace economics will be more nuanced because agents create and capture value differently than apps. The related post on [protocols enabling agent interoperability](/blog/model-context-protocol-mcp) goes further on this point.
**Usage-based pricing** will dominate. You don't pay for the agent. You pay for what it accomplishes. A recruitment agent that sources candidates might charge per qualified candidate delivered. A code review agent might charge per review completed. This aligns incentives. The agent only makes money when it delivers value.
**Performance guarantees** will emerge. Marketplace listings will include SLA-type commitments. "This agent resolves 85% of tier-1 support tickets within 5 minutes with a 92% customer satisfaction rate." If it doesn't hit those numbers, you don't pay. Or you pay less.
**Composability premiums** will matter. Agents that play well with other agents, that expose clean interfaces for orchestration, that follow standard protocols, will command higher prices. Because the real value isn't a single agent. It's a team of agents working together. And agents that integrate smoothly are worth more than brilliant loners.
## Who Wins the Marketplace Race
Three categories of companies are positioning to build agent marketplaces. Each has advantages and blind spots.
**Cloud providers** (AWS, Azure, GCP) have the infrastructure, the enterprise relationships, and the security credibility. But they're historically terrible at developer experience and consumer-grade discovery. Their marketplaces tend to be utilitarian catalogs, not thriving ecosystems.
**AI model providers** (Anthropic, OpenAI, Google DeepMind) have the models, the developer communities, and the technical credibility. But they risk being seen as both platform and competitor. If Anthropic runs the marketplace and also sells its own agents, independent developers worry about competing against the house.
**Vertical platforms** (Salesforce, Shopify, HubSpot) have the domain expertise and the captive user bases. A Salesforce agent marketplace where every agent already understands your CRM data model is compelling. But they're limited to their ecosystem.
My bet: the winning marketplace won't come from any of these categories. It'll come from a company that builds the trust and integration infrastructure as a standalone platform, then lets all the existing players plug in. The equivalent of Stripe for agent commerce. Not a bank, not a retailer. The infrastructure layer that makes the whole thing work.
## The Dark Side
Every marketplace creates perverse incentives. App stores gave us dark patterns, addictive design, and privacy violations at scale. Agent marketplaces will have their own version. It is worth reading about [the broader agentic AI shift](/blog/age-of-agentic-ai-after-chatgpt) alongside this.
Agents optimized for engagement metrics rather than actual outcomes. Agents that create artificial complexity to justify their continued usage. Agents that subtly favor their developer's other products. Agents that collect and monetize data in ways that technically comply with their privacy policy but violate the spirit of user trust.
The marketplace that solves these problems wins. The ones that ignore them will face the same regulatory backlash that hit app stores, just faster and harder. Because an agent that manipulates outcomes is more dangerous than an app that manipulates attention.
## The Timeline
We'll see the first serious agent marketplaces launch in late 2026 or early 2027. They'll be primitive. Limited catalog, manual integration, basic trust mechanisms. Think the App Store in 2008, when the top apps were flashlight utilities and fart soundboards.
By 2028, one or two marketplaces will have established network effects. Developers will build for those platforms first. Enterprises will standardize on them. The winner-take-most dynamics of platform markets will kick in.
By 2030, the leading agent marketplace will be one of the most valuable companies in technology. Not because it builds agents. Because it's the place where agents and organizations find each other, establish trust, and transact.
The App Store didn't build the apps that changed the world. It built the infrastructure that let other people build them. The agent marketplace opportunity is the same, but bigger. Because agents don't just entertain or inform. They work.
And a marketplace for work is a marketplace for everything.