Agent Marketplaces in 2026: App Store 2.0, or a Listing Nobody Installs?

Every large software platform shipped an agent marketplace inside 18 months. Salesforce merged three of its own catalogues into one surface carrying more than 13,600 listings. Not one of these platforms publishes install counts for the agents inside them. That silence is the finding, and the best public proxy available shows 3 agents taking 92% of the actual work across 536 available options.

What an agent marketplace is, and what it is not

An agent marketplace is a catalogue where a platform lists agents, tools and connectors that run inside its own product. Each listing carries a review process and a billing path. It is a distribution surface for software that executes tasks rather than software you open. Discovery is the promise. Procurement is the mechanism that actually works.

That last sentence is the whole argument of this post, so it is worth being precise about why the two get confused. An app store is a place buyers go to look for something. A marketplace listing is a place buyers go to buy something they have already chosen. Those are different products wearing the same interface.

The App Store comparison is doing real damage to how builders plan. It implies a browsing consumer with a credit card and an afternoon. The buyer of an enterprise agent is a procurement function with a security questionnaire and an existing cloud commitment, and that person does not browse.

Three different things get called the same name

The word marketplace currently covers at least three objects with different economics. Conflating them produces most of the bad advice in circulation.

First, the consumer directory, such as the GPT Store and its successor, the app directory inside ChatGPT. Second, the enterprise catalogue attached to a suite you already own, such as AgentExchange or the Agent Store in Microsoft 365 Copilot. Third, the cloud procurement marketplace, meaning AWS Marketplace, Google Cloud Marketplace and Azure Marketplace, where the transaction rather than the discovery is the point.

Only the third has a demonstrated commercial mechanism behind it, and the mechanism has almost nothing to do with browsing. The pattern of platforms absorbing capability that used to be sold separately is covered in more detail in the piece on how agent ecosystems absorb point SaaS.

Every major platform now runs one

The launch cadence between mid 2025 and mid 2026 was close to simultaneous, which is itself a signal. Platforms do not build catalogues because buyers asked for one. They build them because a competitor announced one.

Right-hand column is the point of the table. Every platform publishes supply. None publishes demand. Figures are as reported in the sources listed at the end of this post.

The listings arrived before the buyers

Salesforce reported agent creation on its Slack platform growing 800% year on year, and shipped a $50 million builders initiative to fund partners listing on AgentExchange. Funding the supply side is what a platform does when supply is the constraint it can control and demand is not.

AWS documents the same asymmetry in a quieter way. Its own seller guide states that all new product listings start out with limited visibility, and that sellers must request a visibility change to reach the public catalogue. A default of invisible is a reasonable quality control. It is also a clear statement that shelf space is not what you get on day one.

The number nobody publishes

Here is a test you can run yourself in 10 minutes. Pick any agent marketplace and try to find, for a single listing, how many customers installed it in the last quarter. You will not find it, on any platform, for any listing.

Apple published top charts from the beginning. Google Play shows install bands on every listing. Both had commercial reasons to hide that data and published it anyway, because a marketplace with visible winners is a marketplace buyers trust. Agent marketplaces have chosen the opposite default, in their first 18 months, while telling builders that listing is a distribution strategy.

What the platforms do disclose

They disclose supply, partner logos, funding commitments and customer outcome anecdotes. Salesforce, for instance, reported Humana Dental cutting provider credentialing from 60 days to under 24 hours, and 186,000 customer reviews across the merged catalogue.

Those are real numbers and they answer a different question. A named customer outcome tells you the technology works somewhere. It tells you nothing about whether the marketplace found that customer, and in almost every case of this kind, a field sales team did.

I would treat any marketplace that publishes reviews but withholds installs as answering the question it can win. That is not dishonesty. It is what every catalogue does before the distribution actually exists.

What the usage proxies show

If the platforms will not publish demand data, the honest move is to find the closest measured thing and label it as a substitute. Three sources are worth having.

The clearest usage data comes from outside the marketplaces

Zed, the code editor, published first-party agent telemetry in April 2026 covering 2 million sessions and 15.4 million turns over 90 days. Across that window, 536 distinct agents appeared in the data, and the top 3 accounted for 92% of turns.

Zed's own summary is the sentence to remember: "A long tail exists; it is just not where most of the work is happening." That is not a marketplace. I flag the mismatch again in the limitations section below. It is the largest published dataset on which agents people actually choose to run, and the shape it shows is unambiguous.

The mobile base rate is worse than most builders assume

Sensor Tower measured the top 1% of publishers taking 79% of new worldwide installs in the first half of 2022. That covered the App Store and Google Play. The same 1% took 91% of revenue. The revenue concentration is the number that should worry a builder, because 1,800 publishers out of 184,800 took $42 billion of a $46 billion market.

One qualification matters and is usually left out. Sensor Tower noted the top 1% share had been shrinking since 2019, reaching its lowest point in that period. Concentration in app stores is severe and it is not getting worse, which is a fairer reading than the one usually quoted.

The third proxy is Gartner's vendor census. If roughly 130 vendors are genuinely agentic against thousands claiming it, then a catalogue of 1,000 agent listings contains a large volume of things that will not survive their first production deployment. That holds whatever their install count says.

Distribution through a marketplace is a procurement channel

This is where the realistic assessment lives, and it is more encouraging than the sections above suggest. Cloud marketplaces move real money. Tackle alone reported passing $10 billion of cumulative cloud marketplace transactions by August 2024, having doubled from $5 billion in the preceding nine months.

The reason has nothing to do with a buyer browsing a catalogue. It is that a marketplace purchase draws down against money the customer has already committed to the cloud provider, so the software gets funded from a pre-approved budget instead of a fresh one.

Committed spend is the actual product

Think about what that removes. No new budget line, no separate vendor onboarding, no fresh legal review if the standard contract is used, and a procurement path the buyer's finance team already trusts. A deal that was going to close anyway closes faster and with fewer people saying no.

The fees are consistent with that reading. Google Cloud Marketplace and AWS both charge low single-digit percentages on SaaS transactions, dropping further on large private offers and renewals. A platform charging 3% is running a payment rail. A platform charging 30% is selling you an audience, and no agent marketplace is priced like an audience.

Read the pricing as the platform's own honest disclosure of what it thinks it is providing. It is the most reliable signal in this entire market, and it is free to check.

What a listing is worth, priced honestly

Four things a marketplace could plausibly give a builder. They are worth very different amounts, and lumping them together as distribution is how builders end up disappointed.

The listing is a closing tool, not an opening one. Builders who understand that get real value from it, and they get it late in the sales cycle rather than at the top of the funnel. This is the same discipline the build versus buy analysis for coding agents applies to the purchase side.

Where this argument is weakest

Three genuine problems with everything above, stated plainly because the argument is easier to check that way.

The proxies are borrowed, and that matters

Zed measures a code editor, not a marketplace. Its users are developers choosing agents by reputation, not enterprises buying them through procurement. The concentration it shows could reflect the fact that only 3 agents were any good at coding in early 2026, which is a very different mechanism from marketplace discovery failure.

Sensor Tower measures consumer mobile, where discovery genuinely happens by browsing. Enterprise buying does not work that way, so the base rate may be pessimistic rather than predictive. Neither source is about agent marketplaces, and I would not pretend otherwise.

The honest case for listing anyway

Eighteen months is nothing. AppExchange took years to become a channel and the App Store looked thin in 2009. Judging a distribution surface on its first two years is exactly the error people made about mobile, and the platforms building these catalogues have all seen that history.

There is also a compounding effect that low install counts hide. Being listed when your buyer's procurement team searches for a pre-approved vendor is worth something even if nobody ever browsed to you, and that value grows as more enterprise spend routes through cloud commitments. Salesforce's own data point is instructive here: it reported one partner reaching 80% in-product discovery, which is a real discovery outcome inside a catalogue.

What would change my mind is simple to state. If any major platform publishes an install distribution showing a workable middle class of listings, the discovery case is made. The same goes if a builder with no existing sales motion demonstrates repeatable inbound from a listing alone. Neither has happened as of August 2026.

How to decide whether to list

Three questions, in order. If the first answer is no, the rest do not matter and you should spend the engineering time elsewhere.

Do your existing buyers already hold a commitment with this cloud? If yes, listing converts a budget conversation into a drawdown and pays for itself on one deal. If no, you are building an integration for a procurement path nobody in your pipeline uses.

Is your pricing usage-based? Marketplace metering handles usage billing at a few percent, which is cheaper than building it. If you sell flat annual contracts, the billing benefit largely disappears and only the procurement benefit remains.

Can you name the sales motion that brings buyers to the listing? Write it down before you build. If the answer is that the marketplace will bring them, you have not answered the question. The defensibility question behind the wrapper insult applies here too, because a product that needs the catalogue to find its first customers has a demand problem the catalogue cannot fix.

One technical note that saves real time. If you are listing on a cloud marketplace, support the MCP interoperability standard and Agent2Agent from the start. AWS lets buyers filter by protocol support, and Google requires A2A compatibility for agent listings, so protocol support is now a filter you either pass or vanish from.

Frequently asked questions

Where to start this week

Open your last 20 closed deals and check how many buyers held a committed spend agreement with AWS, Google Cloud or Azure. That single count decides whether a marketplace listing is a finance shortcut for your business or an engineering project with no buyer attached.

Then ask your platform partner manager one question in writing: what is the median number of installs for a listing in this category over the last 90 days? You will probably not get a number. The way the question is deflected will tell you where the catalogue actually is in its life, and it costs you one email to find out.

Related reading