Casper's Notebook
22/11/2026
Field NoteMedium confidence

AI Agent Marketplace Hub

A few weeks ago, I found myself looking at another AI agent directory.

The pitch was familiar. One place to discover agents for sales, support, research, coding, scheduling, finance, and operations. Hundreds of listings. Search filters. Categories. Badges. A promise that buyers would come, compare, and install.

I do not think that is how this market will settle.

My claim is simple.

Marketplaces Will Be Thin

Most AI agent marketplaces will not become important demand hubs. The durable value will sit in workflow distribution, trusted channels, and control layers, not in broad agent catalogs.

That sounds unromantic. It is also what the underlying economics suggest.

An agent is not a ringtone. It is not even a SaaS widget. A useful agent touches live work. It reads tickets, drafts messages, updates records, triggers payments, joins calls, changes calendars, opens claims, or closes leads. The moment it acts, the buyer stops asking “what can it do?” and starts asking harder questions.

What data can it access?

What authority does it have?

How do I monitor it?

How do I stop it?

Who is liable when it is wrong?

Those questions do not get resolved by a marketplace listing page.

They get resolved inside the system where work already happens.

That is why this looks much closer to payments, insurance distribution, and enterprise software procurement than to a consumer app store. I have written before that the biggest moat in AI products will be distribution inside existing workflows, not proprietary models or benchmark wins. I think agent marketplaces are another test of the same idea.

Discovery Is Not Deployment

A catalog can help with discovery. That is real. It is just not enough.

If I am a support leader choosing an AI QA agent, I do not mainly need a menu. I need policy mapping, evidence capture, permissions, audit logs, and integration with the systems of record. That is why a compliance agent behaves less like a downloadable tool and more like a controlled operational system. I made that argument in Call center compliance QA analyzer (collections and insurance). The hard part is not summarization. It is policy-to-evidence.

The same is true elsewhere.

A booking agent for a local salon is not valuable because it can chat nicely. It is valuable because it actually holds the calendar, catches missed calls, confirms appointments, and fails safely. That is a workflow wedge, not a marketplace wedge. I argued that directly in AI voice agent for local service SMBs (missed-call and booking).

In other words, the job to be done is not “find me an agent.” It is “make this part of my operation run with less labor, fewer delays, and acceptable risk.”

Those are different markets.

One is discovery.

The other is adoption.

Marketplaces often overestimate the first and underestimate the second.

Trust Compounds In Channels

There is a more basic first-principles point.

Agents are bought through trust.

Not generic trust in AI. Specific trust in a channel, a workflow, or a platform. Buyers trust the CRM they already use more than a new agent hub. They trust Microsoft inside Microsoft 365. They trust Salesforce inside Salesforce. They trust a vertical software vendor that already handles scheduling, claims, or tickets. They may not love those incumbents. But trust in operations is rarely about affection. It is about acceptable risk.

That pattern shows up well beyond AI.

In insurance, I have argued that distribution and trust matter more than product cleverness. Cover works when it rides on a familiar channel and a credible payment flow, not when it sits alone waiting to be discovered. The same logic applies here, and the analogy is not superficial. See Africa’s insurance gap is a distribution and trust problem, not a product problem.

An AI agent marketplace says: come browse possibilities.

A trusted workflow says: this is already where you work, and here is the narrow authority this automation gets.

The second wins more often.

This is also why I remain skeptical of broad “agent employee” framing. In practice, the deployments that stick are narrower. Fixed scope. Explicit authority. Fast rollback. That is the right shape for real operational adoption, which is why I keep coming back to the autopilot model in Most AI copilots should be replaced by opinionated autopilots with narrow authority and explicit rollback paths.

A marketplace can list an autopilot.

It cannot supply the trust architecture around it.

Payments Show The Pattern

Cross-market comparisons help here.

Look at payments.

India’s UPI became powerful because it was not just a list of payment apps. It was a public interoperability rail with bank connectivity, identity hooks, and enormous acceptance. Brazil’s Pix followed a similar path: instant payments became mainstream because the rail sat under banks and major providers, not because consumers discovered payment tools in a marketplace. China’s super-app model took another route, but the same lesson holds: distribution sat inside apps with daily usage and embedded trust, not in a neutral catalog.

The US and much of Europe are messier. Incumbent card networks, bank fragmentation, and slower account-to-account adoption have preserved more intermediaries and more friction. Yet even there, payment behavior consolidates around trusted endpoints and embedded flows. People do not wake up wanting a “payments marketplace.” They use the option already present in checkout, messaging, banking, or commerce.

Agents will likely follow the same path.

The winning agent layer will often be attached to the place where the data lives and the action happens. CRM. ERP. helpdesk. call center stack. vertical SaaS. productivity suite. maybe telecom in some markets. maybe banking in others.

What can be learned across markets? Interoperability matters. Trust can be designed into rails. Lowering integration friction expands usage. What cannot be copied easily is the installed base. UPI could not be copied by decree into markets with different banking structure, regulation, and smartphone behavior. Pix worked in Brazil because of Brazilian institutional conditions. China’s super-app dynamics rely on a platform and consumer internet history that most markets do not have.

The same caution applies to AI agent hubs. You can copy the front-end pattern of a directory. You cannot copy the installed workflow, data gravity, and trust relationships that make agents actually deploy.

The Real Hub Is Control

If there is an “agent marketplace hub” worth building, it probably does not look like a mall.

It looks like control infrastructure.

Identity and permissions.

Audit trails.

Evaluation.

Usage policy.

Fallbacks.

Model routing.

Cost controls.

Memory.

Observability.

Procurement and billing.

That is less glamorous than a storefront. It is closer to what enterprises and serious SMBs will actually pay for.

I have argued before that AI agents win on persistent memory and context, not model capability. I would extend that here: agent ecosystems will be governed by persistent control and context, not by catalog depth.

This is where some platforms may still use marketplace language, but the business underneath will be different. The “marketplace” becomes a surface attached to an operating system for agents. Buyers may browse there. The money and defensibility come from governance and workflow embedding.

That is a much tougher business to build.

It is also more real.

Consumer Might Differ

There is one place where I would soften the argument.

Consumer-facing agent discovery may matter more than enterprise discovery.

If personal AI assistants become common, a consumer may indeed browse a set of plug-ins, skills, or task-specific agents. But even there, I suspect usage will concentrate around a few trusted defaults and a few embedded contexts. Consumers say they want choice. Then they use the default keyboard, the default map, and the payment option already attached to checkout. Arsenal supporters say they want beautiful complexity too, until the match turns and they beg for one reliable passing lane out of pressure. Systems under stress reveal what people really value.

Enterprise buying reveals it even faster.

My Confidence

Medium.

I am confident that broad, neutral agent marketplaces will struggle to become durable demand aggregators on their own.

I am less confident about the exact boundary between “marketplace,” “platform,” and “control plane,” because the winners may package all three together. Some hubs will attract supply. Some will generate leads. Some will matter for developers. I just do not think the catalog itself is where enduring power sits.

I am also open to a consumer exception, especially if a dominant personal assistant creates enough trust and standardization that third-party agents can be installed with clear permissions and easy rollback.

What Would Change My Mind

A real test would be this:

If a broad agent marketplace, without owning a major workflow surface, can repeatedly drive high-retention deployments of third-party agents into businesses across multiple categories, then I am wrong.

Not trials. Not installs. Not curiosity traffic.

Retained production usage.

I would want to see evidence that buyers discover agents in the marketplace, deploy them into live workflows, keep them running for [[clear: period to verify, perhaps 6-12 months]], and expand spend without the marketplace first becoming the control plane or being bundled into an incumbent workflow platform.

A second test is concentration.

If the majority of meaningful agent purchasing starts with neutral hubs rather than with incumbents like Microsoft, Salesforce, ServiceNow, Shopify, HubSpot, industry software vendors, or communication platforms, then the marketplace thesis gets much stronger.

For now, I think the opposite is more likely.

The center of gravity in agents will not be where they are displayed. It will be where they are trusted to act.

Sources

  • https://web.archive.org/web/20261118101434/https://www.npci.org.in/what-we-do/upi/product-overview
  • https://www.bcb.gov.br/en/financialstability/pix_en
  • https://www.salesforce.com/agentforce/
  • https://openai.com/chatgpt/team/