App Stores for Agents: The New Distribution Battle

Table of Contents

• Why an App Store for Agents Is Even a Thing
• The Contenders: Who Wants to Own the Store
• Foundation-Model Providers
• Cloud Hyperscalers
• Incumbent SaaS Platforms
• The Independent Marketplaces
• Why the App Store Analogy Breaks Down
• The Real Battleground: Trust, Billing, and Defaults
• What Discovery Looks Like When the Buyer Is an Agent
• The Economics: How the Rake Gets Rewritten
• What Agent Builders Should Do Right Now
• Insights Most People Overlook
• References

Why an App Store for Agents Is Even a Thing

Every platform shift produces a distribution land grab, and the people who win it rarely build the best product. They build the best storefront. The web had Yahoo's directory and then Google. Mobile had Apple's App Store and Google Play. SaaS had AppExchange, the Atlassian Marketplace, and a hundred lesser "integration directories." Now agentic AI-as-a-service is reaching the point where there are too many agents for a buyer to find, evaluate, and trust on their own, and a familiar instinct kicks in: build the store.

The pressure is real. A company that wants an agent to handle accounts-payable reconciliation, or to triage inbound support tickets, or to run outbound SDR sequences, faces a fragmented mess. There are venture-backed vertical agents, open-source agents, agents bolted onto SaaS suites the company already owns, and agents that are really just a prompt and a wrapper. The buyer has no neutral way to compare them on the thing that actually matters, which is whether the agent reliably completes the task at an acceptable cost per outcome.

That gap is the opening. Whoever fills it becomes the place agents get distributed, and distribution, as the incumbents keep reminding everyone, beats capability. The catch is that nobody has agreed on what an agent store should even be.

The Contenders: Who Wants to Own the Store

There are four camps, and they want incompatible things.

Foundation-Model Providers

OpenAI's GPT Store was the loudest opening move, and it's instructive precisely because it underdelivered. Launched with fanfare in early 2024, it demonstrated that listing tens of thousands of "GPTs" is easy and that getting anyone to pay for them, or even find the good ones, is hard. The lesson the model providers took away is that the store isn't a directory problem; it's a trust and monetization problem. Anthropic's push around the Model Context Protocol points at a different theory of distribution, in which the valuable real estate isn't a consumer-facing catalog but the standardized connective tissue that lets any agent reach any tool or data source. Own the protocol, and you sit underneath every store rather than competing to be one.

Model providers have a structural advantage here: they're already in the inference path. Every agent call runs through them. That gives them billing data, usage telemetry, and the ability to make their own agents the default. It also gives them the platform-risk problem from the agent builder's side, which is its own cluster topic worth understanding before you build on anyone's foundation.

Cloud Hyperscalers

AWS, Microsoft, and Google already run the marketplaces where enterprises buy software with committed cloud spend, and those marketplaces are quietly the most powerful distribution channel in B2B software. The reason is boring and decisive: procurement. When an agent can be purchased through AWS Marketplace and drawn down against an existing enterprise discount program commitment, it skips the entire vendor-onboarding gauntlet. No new MSA, no new security review from scratch, no new payment relationship. Microsoft's Copilot ecosystem and its agent builder tooling lean on the same advantage from inside the Office and Azure footprint most enterprises can't avoid.

Hyperscalers don't need to win on agent quality. They need to win on the friction of buying, and they're already ahead there.

Incumbent SaaS Platforms

Salesforce (AgentForce), ServiceNow, HubSpot, and the rest are converting their existing app marketplaces into agent marketplaces. Their pitch is gravity: the agent runs where your system of record already lives, with permissions and data access pre-wired. This is the same incumbent data-moat argument that shows up across this beat, and it's genuinely strong for agents that need deep, governed access to enterprise data. The risk for these vendors is that hosting third-party agents can accelerate their own unbundling, a tension explored in the SaaS-suite-unbundling discussion. A marketplace that helps customers replace your seats is a strange thing to build, and several incumbents know it.

The Independent Marketplaces

Then there are the neutral aggregators and agent registries trying to be the Switzerland of distribution, plus the open-agent directories emerging around MCP servers and agent registries. Their bet is that buyers will distrust any store run by a vendor with skin in the game, and will want a neutral place to compare agents on verified performance. It's a good thesis with a brutal go-to-market: neutral marketplaces have no inference path, no procurement hook, and no installed base. They have to manufacture trust from scratch, and trust is the whole game.

Why the App Store Analogy Breaks Down

It's tempting to assume agent stores will rhyme with the mobile app stores, complete with a 30% platform tax, an approval process, and ranking algorithms gamed by everyone. That mental model will mislead you, because three things are fundamentally different.

First, the unit of value is an outcome, not a download. You don't "install" an SDR agent and forget about it. You pay it per qualified meeting, or per resolved ticket, or per reconciled invoice. The store, if it's doing its job, has to measure and attest to those outcomes. That's a far heavier lift than hosting a binary and counting installs, and it's why per-outcome pricing reshapes everything downstream.

Second, the buyer might not be human. When an orchestrator agent selects a sub-agent to handle a subtask, the "discovery" happens machine-to-machine, in milliseconds, with no marketing landing page involved. App store optimization as we know it, screenshots and review counts and keyword-stuffed titles, is irrelevant to a buyer that reads a structured capability manifest and a verified reliability score.

Third, the lock-in mechanics invert. Mobile apps locked in users through data and habit on the device. Agents create lock-in through accumulated context, learned workflows, and the "agent of record" relationship with a given business process. The store that captures that relationship captures something stickier than any app install ever was, which is exactly why the agent-of-record concept is becoming a strategic flashpoint.

So the analogy gives you the right intuition, that distribution will consolidate, and the wrong details, that it'll look like the App Store. It'll look like something closer to a labor marketplace fused with an API gateway and a credit bureau.

The Real Battleground: Trust, Billing, and Defaults

Strip away the storefront and three control points decide who wins.

Trust and verification. A buyer can't try forty agents. They need a credible signal that an agent does what it claims, safely, without leaking data or going off-script. Whoever owns the verification layer, the equivalent of a credit score for agent reliability and security, becomes the gatekeeper everyone has to go through. This is where a16z and others have argued the durable value of the emerging AI agent infrastructure stack actually sits, not in the agents themselves but in the trust and observability scaffolding around them.

Billing. Whoever owns the payment relationship owns the customer, full stop. This is the lesson every SaaS marketplace already learned. If the store handles metering, invoicing, and settlement, the agent builder becomes a supplier and the store becomes the brand the customer trusts. Hyperscalers and incumbents understand this in their bones, which is why their marketplaces are billing rails first and catalogs second.

Defaults. The most undervalued asset is being the default. When Microsoft 365 ships with a built-in agent for a task, the third-party agent that does it 15% better still loses to the one that's already there and already paid for. Defaults are how incumbents convert distribution into a moat that capability can't easily breach, a dynamic that runs through the entire incumbents-versus-startups question.

What Discovery Looks Like When the Buyer Is an Agent

Here's the part most coverage misses. The interesting agent discovery isn't a human browsing a catalog. It's an orchestration layer dynamically selecting which specialized agent to invoke for a subtask, then paying for it, then moving on.

That implies a discovery infrastructure that looks nothing like an app store. It looks like a registry of machine-readable capability descriptions, a reputation system that other agents can query, a pricing API so the orchestrator can cost-optimize across providers, and a standardized invocation interface so swapping agent B for agent A is a config change, not an integration project. Protocols like MCP and the various agent-to-agent communication efforts are the early plumbing for exactly this. Whoever standardizes that plumbing shapes the market more than whoever builds the prettiest marketplace front end, because the front end is increasingly an afterthought.

This also means the commoditization risk for agent builders is severe. If your agent is one interchangeable option an orchestrator picks on price and reliability, you have no brand, no relationship, and no pricing power, only a leaderboard position. The agents that escape that trap will be the ones with proprietary data access or deep domain integration the orchestrator can't route around.

The Economics: How the Rake Gets Rewritten

The 30% app store tax was possible because Apple controlled the only door to a billion phones. No agent store controls a comparable chokepoint, at least not yet, so the economics will look different.

Expect the take rate to compress and to fragment by control point. A hyperscaler marketplace that provides billing, procurement, and committed-spend drawdown can justify a meaningful cut because it removes enormous friction; enterprises will happily route an agent purchase through AWS to avoid a six-month procurement cycle. A neutral directory that provides only discovery can justify almost nothing, because discovery alone is a commodity. The fat margins accrue to whoever bundles the most control points, trust plus billing plus defaults, into one relationship.

There's a deeper shift underneath all of this. Agent spending increasingly comes out of a labor or operations budget, not a software budget, which changes who the buyer is and how they evaluate cost. McKinsey's work on the economic potential of generative AI frames the prize in terms of labor-equivalent value, and that framing matters for stores: a marketplace that lets a CFO buy "outcomes" charged against an opex labor line, with clean attribution and audit trails, is selling into a far bigger budget than the IT software line ever held. The store that speaks the language of labor procurement, not software licensing, captures that reframing.

What Agent Builders Should Do Right Now

If you're building a vertical agent, the distribution battle is not abstract; it determines whether you have a business.

Don't bet your distribution on a single store. The platform risk of building on one provider's marketplace is real, and the provider can launch a competing first-party agent the moment your category looks lucrative. List where the procurement friction is lowest for your buyer, which today usually means the hyperscaler marketplace your enterprise customers already buy through.

Invest early in making your agent legible to machine buyers. A clean capability manifest, transparent pricing, and verifiable reliability metrics will matter more over time than any human-facing landing page, because more of your "buyers" will be orchestration layers.

And protect the relationship. The single most valuable thing you can own is being the agent of record for a specific business process, with the accumulated context and trust that implies. That relationship is what no store can easily take from you, and it's the only durable defense against being commoditized into a line item on someone else's leaderboard.

Insights Most People Overlook

• The store that wins may never show a human a catalog. Everyone is designing storefronts for human browsers, but the highest-volume agent purchases will be machine-to-machine, selected by orchestrators in real time. The winning "store" might be a pricing-and-reputation API with no consumer UI at all, which means most of the current marketplace investment is aimed at the wrong buyer.

• Verification, not discovery, is the real moat. Listing agents is trivial; attesting that they reliably and safely do the job is hard, expensive, and defensible. The entity that becomes the credit bureau for agent reliability holds more power than any catalog owner, and almost nobody is building that yet.

• Incumbent SaaS marketplaces are quietly self-cannibalizing. A vendor that opens its marketplace to third-party agents is helping customers find tools that replace its own seats. The most strategically coherent move, refusing to host agents that erode your seat revenue, is also the one that makes you look closed and behind. That contradiction will fracture incumbent strategy over the next two years.

• The 30% rake is dead on arrival, and that's bad for builders too. A compressed, fragmented take rate sounds like good news for agent makers. But it also signals that no store can offer the demand-generation Apple once did, which means builders inherit the burden of creating their own demand. Cheaper distribution that delivers no customers is not a bargain.

• "Default beats better" is more lethal for agents than it was for apps. Because agents are bought on outcomes and switching requires re-establishing trust and context, an inferior pre-installed agent enjoys even more inertia than a pre-installed app did. The bundled first-party agent doesn't need to be good. It needs to be already there and already paid for.

References

• Introduction to the Model Context Protocol
• Andreessen Horowitz: The AI Agent Infrastructure Stack
• McKinsey: The Economic Potential of Generative AI