AI Agents on AWS Marketplace: A Seller's Guide

AWS Marketplace has a dedicated category for AI agents and agentic tools, two supported deployment options, and a quality bar you have to attest to. The listing mechanics will feel familiar if you have sold SaaS there before. The part that is genuinely new is deciding what a unit of your agent’s work is worth — and you have to decide that before you publish, because the pricing model freezes.

Most of the questions we hear about agent listings are really questions about metering. The listing form is the easy half. An agent that reasons, calls tools, and works for minutes at a time does not have an obvious billing unit, and marketplaces make you commit to one early.

Here is what AWS actually supports today, what Microsoft has alongside it, and the decisions worth making before you open the listing form.

What is an AI agent product on AWS Marketplace?

An AI agent product is a listing for software that reasons and acts on a user’s behalf rather than following fixed rules. AWS defines agents as “software systems that use artificial intelligence to reason, plan, and complete tasks on behalf of humans or systems,” and distinguishes them from traditional software because they “operate independently, adapting through multi-step processes to achieve specific goals.”

The category covers two related things:

Agents — the systems that combine a reasoning model with tools to do work end to end.
Agentic tools — the components that make agents more capable. AWS names knowledge bases, guardrails, integration protocols including MCP servers, specialized APIs and microservices, and business logic components.

AWS groups the products it expects into content creation, data analysis, customer service, business process automation, security and compliance, and developer tooling — explicitly “not an exhaustive list.”

There is a quality gate attached. If you select an AI agent tool type when listing, you are asked to confirm that your agent uses reasoning models and demonstrates autonomous capabilities. That attestation exists to keep the category from filling with anything that calls a model once.

The two deployment options, and how to choose

AWS supports API deployment and container deployment for agent products, and the choice is mostly about whose environment the data sits in.

	API deployment	Container deployment
Hosting	Vendor-hosted endpoints	Customer’s own AWS environment
Data control	Processed on vendor servers	Greater customer control
Infrastructure	Minimal — uses your infrastructure	Customer manages it
Customization	Limited to API capabilities	Full control over the environment
Updates	Automatic, on your schedule	Manual, on the customer’s schedule
Regulatory fit	Bounded by your certifications	Easier to adapt to specific requirements

AWS’s own guidance on choosing is short and worth taking at face value. If your buyers work with data that cannot leave their environment, container deployment is likely the answer. If your model is large or needs specialized hardware, API deployment is more practical. And consider the operational overhead of maintaining both, because supporting two delivery paths is a standing tax on a small team.

The honest version for most ISVs: pick the one your product already is. An agent that runs as a hosted service should be listed as one. Repackaging it into a container to look more enterprise-ready creates a second product to support and a second support surface to staff.

The pricing decision you cannot defer

This is the part that catches teams out. AWS supports four SaaS pricing models — subscriptions, contracts, contracts with pay-as-you-go, and free — and states plainly: “Once you create your listing and publish it to limited, you can’t change the pricing model.”

That means the modelling has to happen before the listing does. Three questions decide it:

What is the unit a buyer would recognise? Seats work when a human sits behind each agent session. They stop working when one person’s agent does the work of a team. Task, run, resolved ticket, or document processed all describe the value better, and all require you to define them precisely enough to defend in an invoice dispute.

Is your cost per unit stable? An agent’s cost varies with how hard the task is — tokens consumed, tools called, retries. If a hard task costs ten times an easy one and you price a flat per-task rate, your margin moves with the customer’s use case rather than with your pricing.

Does the buyer need a predictable number? Enterprise procurement does. This is why contract pricing with pay-as-you-go on top is the shape a lot of agent products settle on: a committed floor the buyer can budget, with metered overage for the variable part. Contract pricing vs usage pricing on AWS Marketplace covers what each model does to your metering obligations.

A useful discipline is to price the outcome one level above the mechanism. Buyers do not want to buy tokens, and pricing on tokens hands them a reason to optimise your revenue down. They also do not want a flat seat price that makes heavy use free. The unit that survives is usually the one your customer already counts in their own operating reports.

What the other marketplaces are doing

Microsoft Marketplace has an AI Apps and Agents category alongside its existing offer types, and it is treated as a first-class offer type — Microsoft lists it among the types eligible for the Standard Contract for Microsoft Marketplace, next to Azure Applications, containers, virtual machines and SaaS.

Microsoft’s pricing rules are stricter than AWS’s in one respect worth knowing before you list on both: a SaaS offer uses either flat rate or per user, all plans in one offer must share the model, and “after you publish your offer, you can’t change the pricing model.” If you plan a consumption-shaped agent there, that means flat rate with metered dimensions layered on, not a per-user plan you intend to convert later.

The practical consequence of listing across marketplaces is that your metering has to speak more than one dialect while your product counts one thing. Decide the internal unit once, then map it outward.

What to settle before you open the listing form

The unit. Written down, with an example invoice line a buyer would accept.
The deployment option, and whether you are genuinely prepared to support two.
The contract path. Standard contract or your own EULA — EULAs and custom terms in marketplace deals covers what each costs you in cycle time.
What you meter and when. Agents fail, retry, and partially complete. Decide now whether a failed run bills.
What the buyer sees at the end of a month. If your usage report cannot be reconciled against the marketplace invoice by a finance analyst who has never met you, expect disputes.

Frequently asked questions

What is the AI Agents and Tools category on AWS Marketplace? It is the AWS Marketplace category for agent products and the agentic components that support them — knowledge bases, guardrails, MCP servers, specialized APIs, and business logic. Sellers choose one to three categories when listing.

Can I list an MCP server on AWS Marketplace? Yes. AWS names integration protocols including MCP servers among the agentic tools suited to the marketplace, alongside knowledge bases and guardrails.

Should I deploy my agent as an API or a container? API deployment if your model needs specialized infrastructure or you want to keep it in your environment. Container deployment if buyers need the data to stay in their own AWS account.

Can I change my agent’s pricing model after listing? No. AWS states that once you publish a listing to limited, you can’t change the pricing model. Microsoft applies the same rule to SaaS offers after publication.

How should an AI agent be priced on a marketplace? Most agent products end up on a committed contract with metered overage: a floor procurement can budget, plus usage for the variable part. Price the outcome, not the tokens.

Does Microsoft have an equivalent category? Yes — Microsoft Marketplace has an AI Apps and Agents category, and it is one of the offer types eligible for the Standard Contract for Microsoft Marketplace.

Takeaways

AWS defines an AI agent as software that reasons, plans and acts independently, and gives agents and agentic tools their own category.
Two deployment options are supported: vendor-hosted API endpoints, or containers running in the customer’s own AWS environment.
Selecting an agent tool type requires attesting that the product uses reasoning models and shows autonomous capability.
The pricing model freezes at limited publication on AWS and at publication on Microsoft. Model it before you list, not after.
Seats price the wrong thing for an agent. A recognisable unit of work, with a committed floor and metered overage, is the shape most agent products land on.
Microsoft has an AI Apps and Agents category, so a cross-marketplace agent needs one internal unit mapped to two sets of rules.

Agent products make the same demand of your systems that any usage-priced listing does, only with a less obvious unit: metering that is correct, and entitlements your product can check before it does expensive work. See how Suger handles AI agent and MCP server distribution across the marketplaces you sell on.