Agent Marketplaces and App Stores: Buy or Build

The agent store has arrived

Your CRM now has an agent marketplace. So does your helpdesk, your productivity suite and your ecommerce platform, and the large AI providers run their own directories of agents and connectors. A sales operations manager can install a prospect research agent on Monday morning and have it working by lunch, and the procurement team will find out about it on the next invoice.

That convenience is real, and for many common tasks buying is plainly the right decision. It also brings the familiar software buying questions back, sharper than before, because an agent does more than display data. It acts on it.

What you are actually buying

A marketplace agent is usually a bundle of four things, and each deserves a separate look:

• The instructions and workflow the vendor has designed for a task, such as lead research or ticket triage
• Connectors that give it access to your systems, often through the host platform's permissions or MCP servers
• A model, chosen by the vendor, which may change without notice
• A pricing model, increasingly per task, per outcome or per credit rather than per seat

The workflow is the part you can see in a demo. The connectors and permissions are the part that matters for risk, and the pricing is the part that matters a year later.

Buy vs build for AI agents

Our general thinking on this is in build vs buy for AI features. Agents shift the balance slightly towards buying for generic work, because vendors can spread development costs across thousands of customers, and slightly towards building for anything that touches sensitive data across systems.

Questions to ask before installing a marketplace agent

1. What exactly can it read and change, and can those permissions be narrowed?
2. Where is our data processed and stored, and is it used to train anything?
3. Which model does it use, and will we be told when that changes?
4. What does it log, and can we see every action it took?
5. What happens when it is wrong: can actions be reviewed before they take effect?
6. How is it priced at ten times our current volume?
7. Who built it: the platform vendor, a verified partner or an unknown developer?
8. How do we export our data and configuration if we leave?

Question seven matters more with agents than with ordinary apps. Third-party agents and connectors in open directories vary widely in quality, and a poorly built connector with broad permissions is a genuine security risk. Our post on MCP server security explains what to look for. For general vendor questions, see what to ask AI software vendors.

Pricing traps with bought agents

• Outcome pricing with loose definitions. Paying per resolved ticket is appealing until you learn what counts as resolved.
• Credits that expire or tiers that jump sharply at the next usage band.
• Platform upgrades required to access the marketplace at all.
• Several overlapping agents installed by different teams, each billing separately for similar work.

An illustration: an agent charging a small fee per resolved support conversation looks cheap at 500 conversations a month. At 8,000 a month it may cost more than building and running your own, and the vendor's definition of resolved will determine whether it is fair. Do the arithmetic at your expected volume, not the pilot's.

The case for building, stated fairly

Building costs more upfront and takes longer. You own the maintenance, the monitoring and the evaluation. In return you get an agent designed around your process rather than a generic one, access controls you define, a model you choose and can change, and costs that scale with infrastructure rather than a vendor's price list.

Building is also the only option when the agent needs to work across your own systems in ways no marketplace vendor anticipated. That is typically the agent that handles your core operational process, the one where a generic version would force your team to work the vendor's way.

A practical mix

Most businesses we speak to end up with a sensible portfolio: bought agents for meeting notes, generic research, basic support deflection and writing help, and one or two built agents for the processes that matter most. The discipline is keeping a register of which agents are installed, what they can access, and what each costs, before the list grows by itself.

At SpiderHunts we are happy to tell a client to buy the marketplace agent when it fits, and we do so often. When it does not, we build through our AI agent development service, usually connecting to the same platforms rather than replacing them. The broader effect on your software choices is covered in what agentic AI means for software buying.

Frequently asked questions