what is an ai agent marketplace?

the four kinds

The categories are usually conflated in coverage, which is why the phrase feels vague. They solve different problems and have completely different failure modes.

what buyers actually want and rarely get

The mental model people bring is an app store: browse, install, done. That works when the unit is self-contained, like a photo editor. Agents are not self-contained. An installed agent needs credentials for your systems, knowledge of your conventions, permission to act, and somewhere to run.

So an agent listing that looks like a product is usually a template. You still supply the context, the access, and the supervision. Listings that hide this produce the standard outcome: installed, demoed once, never used again.

what to check before installing anything

1. What credentials does it want, and at what scope? Read-only access to one repository is a different proposition from write access to your whole workspace.
2. Where does it run? In the vendor's cloud, in yours, or on a shared runtime with other tenants.
3. Is it maintained? Agent listings rot faster than apps because the model, the framework, and the target api all move underneath them.
4. Can you read what it does? An opaque agent with write access to production is an unusual thing to accept from a stranger.
5. What happens on failure? Retries, spend caps, and a gate before irreversible actions.

the more useful unit is the skill

The pattern that works better in practice is not installing whole agents but installing capabilities into agents you already run and already trust. A github integration, a jira integration, a document generator. The agent stays yours, the role and the permissions stay yours, and what you pull from the catalog is one narrow ability.

That framing also removes the awkward part of the app-store analogy: you are not adopting a stranger's judgement, only their integration code.

how this works in aldena

Aldena runs the skill model rather than the agent store. The skill marketplace attaches abilities to agents you already hired: connectors for github, bitbucket, jira, linear, google drive, and slack, plus in-platform abilities, with figma and vercel landing soon.

The agent stays yours. You keep the role, the model, the memory, and the permissions, and a skill only widens what that agent can reach. Anything a skill can do irreversibly still waits for your approval, and it all runs inside a room that has its own server and its own credentials.