BusinessNew Release11 min readPublished

AWS Marketplace Now Sells API Calls Directly to AI Agents

A research agent compared two listings and paid one cent for a search. Its developer decided how much it could spend, which sellers it could use and whose AWS bill would carry the charge.

DA
Digital Applied
PublishedNovember 21, 2026
Read time11 min
SourcesAWS launch · seller guide · sample code

An AWS research agent needed fresh information, looked at two Marketplace listings and paid Tavily $0.01 for one web search. No one had to open a Tavily account to finish the task. On November 19, 2026, AWS added its own billing system to Amazon Bedrock AgentCore payments so a customer can let an agent buy individual requests from participating sellers in US dollars.

The small charge hides the more consequential event. An agent can now make a first purchase from a seller on behalf of the AWS account that owns its payment manager. A team enabling this has three decisions to make before the first search runs. Which products may answer this question? What is one answer worth? Who can tell later which task bought it? AWS Marketplace launch announcement

Key takeaways
  1. 01
    An eligible listing quotes each requestA participating seller returns a signed, fixed US-dollar quote through x402. The AWS Billing connector lets AgentCore payments process it against an existing AWS account.
  2. 02
    The example paid for one searchAWS reports one $0.01 transaction after its agent compared two listings. The price belongs to that request. Discovery steps and model usage have different roles in the cost of the complete task.
  3. 03
    The two spending limits live in different placesIn AWS’s sample, AgentCore enforces the $0.05 session budget. Buyer-side code rejects a quote above $0.02 per request. The purchasing role cannot create a higher-budget session.
  4. 04
    The bill and task log answer different questionsMarketplace meters product charges to the account that owns the payment manager. Keep the product and payment IDs beside the task outcome so finance and engineering can reconcile a purchase.

01 — The purchaseWhat does one cent buy?

The AWS example starts with a question about that week’s agent-payment news. Fresh information matters because the model cannot know what appeared after its training. The agent searches the Marketplace catalog, inspects listings and obtains the address of a seller’s paid endpoint. AWS reports that it compared two listings, then bought a single search from Tavily. The agent received search results it could use to answer the original question with sources.

A search request is a useful first unit because the agent can recognize both the need and the response. It can state what it asked and show which results helped answer the question. A buyer might also authorize a single lookup from a data provider. In both cases the paid item needs a boundary. If a listing delivers one search response, exploring a second query can mean a second purchase even when the human regards both as one research task.

That one-cent figure is evidence of a purchase in AWS’s published run. A planning estimate also needs the frequency of follow-up queries and the cost of running the agent. For a recurring report about familiar sources, an existing feed might already answer the question. Give the agent a reason to buy a fresh result, such as a missing current document, and a reason to stop after it has one. AWS’s worked Tavily purchase

02 — From quote to dataThe seller names a fixed price

The initial request reaches a product that supports the x402 protocol. Its endpoint responds with HTTP 402, Payment Required, and includes a signed quote. Think of the difference between a price on a shelf and a price presented at checkout. The listing can show a typical price or a range. The signed quote specifies the amount for the request actually being made. The listing estimate helps the agent choose where to look.

Consider a listing that describes a normal lookup as costing roughly a cent. A complex request might receive a different exact quote. The difference matters even if the developer approved that seller yesterday. Have the buyer present the quoted amount to its purchase check before authorizing it. A catalog cache is useful for deciding which seller to ask first. Its stored estimate gives the team little recourse if the transaction record later shows that the agent accepted a higher price.

The buyer passes that quote to AgentCore payments using the AWS Billing connector. Marketplace verifies it using the key registered for the seller’s product. The purchasing path checks eligibility and the session budget. If the purchase is authorized, the agent receives a payment token. It presents that token when repeating the request, and the seller verifies it before supplying the content. A refusal because the session has insufficient funds produces no charge for that refused payment.

The two requests to the seller serve different purposes. A monitoring system that simply counts outbound web requests might record two attempts around a single paid response, in addition to the earlier discovery calls. Use the payment record to count purchases and the API response to record deliveries. That distinction also helps explain a timeout after payment, when the system needs to inspect what happened before buying again. How x402 represents a paid request

The offer at checkout

The quoted amount is fixed for that request under the x402 exact scheme AWS Billing supported at launch. A different query, product or later quote can have a different price. Read each quote against the limits for the current task.

03 — Spending authorityWho owns the five cents?

The developer’s setup creates a payment manager with an AWS Billing connector. The AWS account also needs the Marketplace service-linked role used for these purchases. For each task, the application can create a time-limited payment session with its own budget and hand its identifier to the agent. AWS’s example gives one role permission to create the session and a different role permission to make purchases within it.

Read the two prices in the sample carefully. Its session begins with $0.05, which AgentCore enforces on the service side. Its buyer code also sets a maximum quote of $0.02 for a single call. That second check happens in the client. A team building a different buyer must preserve the per-request check in any path that can submit a purchase if it wants the same rule. Telling the model that two cents sounds reasonable does not implement it. AWS’s buyer sample and its two limits

Why give the roles different permissions? Suppose a task gets four cents into an inconclusive search and wants to keep trying. It can use the authority left in its current session if another purchase fits the rules. Obtaining a new five-cent session belongs to the application that holds the budgeting role. That gives a manager a concrete option when the research is worth more money. A new allowance can be recorded as a deliberate decision, tied to the same business task.

Here is the arithmetic for the published limits. Only the first purchase below happened in the example. The other rows are independent scenarios using the same five-cent starting budget. They show why the per-request ceiling and the remaining session balance answer different questions.

US dollars. The first row is the AWS walkthrough. The other rows illustrate budget decisions, assuming product and seller approval.
Available before requestNew seller quoteDecision under these limits
$0.05 at session start$0.01Allowed in the reported run. The search leaves $0.04.
$0.05 at session start$0.03The buyer rejects the individual quote although the session has enough money.
$0.01 remaining$0.02The quote meets the per-call cap. The remaining session budget cannot cover it.
$0.01 remaining$0.01Within both limits. A successful payment uses the last cent.

04 — Supplier controlChoose who may answer the question

An agent with a low spending cap can still send a commercially sensitive question to an unsuitable provider. A team approving a search service needs to consider what text the agent may send and what the service may do with it. A public product launch and an unpublished acquisition target could fit the same one-cent budget. They deserve different data-handling decisions before the search tool sees either query.

AWS describes approved-seller rules through AgentCore policy. An organization using AWS Private Marketplace continues to control the products its agents may buy. Those checks answer different questions from the dollar limit. A cheap quote can still be ineligible. Agree on a short list of approved products for the first workflow and have the application explain a policy refusal without passing the original question to a new provider.

The Marketplace connector in AgentCore Gateway is an optional discovery tool. It helps an agent search listings and obtain eligible endpoints. A team that has chosen a single supplier can use its configured endpoint directly, as the AWS example allows. The AWS Billing connector in the payment manager still handles payment in either setup. This matters when an organization wants one approved data source today and the ability to compare more sources later.

One licensed source
Configure the approved endpoint and let the task buy a lookup only when the stored information is stale.
Fixed supplier
Several suitable products
Use Marketplace discovery, restrict eligible products and require a quote within the task’s purchase limits.
Limited choice
Sensitive or unapproved query
Keep the query inside approved systems. Ask the task owner to choose another authorized source where necessary.
No purchase

05 — ReconciliationWhere does a one-cent purchase appear?

The charge belongs to the AWS account that owns the payment manager. AWS explains that the first purchase from a seller creates a Marketplace agreement for that account. Marketplace then meters charges and each purchased product appears as a Marketplace line item on the bill. Settlement is asynchronous. An invoice view immediately after a run can lag behind the successful payment response.

Finance can use AWS Cost Explorer to examine spending and AWS CloudTrail to investigate the activity. The sample also prints the amount, product ID and payment ID. Save those identifiers beside the application’s task ID and its result. A product line on the monthly bill tells finance which service accumulated charges. It will take the application record to explain that the third search in a given session was needed to verify a conflicting date.

Suppose a team notices that a frequently purchased result rarely makes it into the final answer. A ledger showing only successful payments makes that look like healthy usage. Pairing payment evidence with the result lets the team see purchases that delivered technically valid data but little benefit. Keep payment tokens out of that record. A token is a credential the seller can accept, while a payment identifier lets an operator refer to the transaction. AWS’s payment helper and logging guidance

A failure after the payment check deserves its own outcome in that log. A timeout might mean the seller sent data that never reached the client. It might also mean the service was unavailable. In either case, capture the prior payment reference before initiating another paid lookup. The team can investigate the original attempt or obtain approval for a replacement while keeping an accurate account of money spent on an answer that the user never received.

06 — Task economicsWhat happens at 10,000 tasks?

For a hypothetical month of 10,000 tasks, one one-cent search on each task would produce $100 in seller charges. If every task instead used its entire five-cent allowance, those charges would reach $500. If only a quarter of the tasks needed new information and each of those bought one one-cent search, the total would be $25. None of these figures includes model inference, runtime or any other cloud service used to complete the task.

The assumptions matter more than the multiplication. The $500 ceiling holds only if each of those 10,000 tasks maps to one five-cent session. An application that opens a fresh session whenever a task resumes after an interruption changes the arithmetic. Two attempts might each receive authority to spend. Persist the task-to-session mapping and decide how a human-approved increase is recorded before using the per-task ceiling in a forecast.

The same calculation helps compare a standing contract with per-request buying. At an assumed one cent per usable query, a hypothetical $40 monthly feed costs the same as 4,000 individually purchased lookups. That break-even assumes both products return equally usable answers and grant the same rights. If the feed permits local reuse while the per-call product requires a new purchase for every task, the query count alone obscures an important difference. Estimate distinct lookups after permitted reuse before pricing either route.

A useful weekly measure is seller spend divided by completed tasks that met the acceptance criterion. In a research product that could mean an answer with a current, relevant source. Track refusals and abandoned tasks alongside successful ones. A rule that cuts spending by rejecting most legitimate searches can make the budget look good while the product gets worse. How to budget long-running agent tasks

07 — Selling a requestCan a publisher sell one answer?

Consider a publisher with a specialized events database. Its prospective customer may need the next date for one conference and have little use for a monthly subscription. Selling one response through Marketplace could serve that request if the product can quote a definite price, return useful data promptly and give the buyer clear terms for using it. The ability to collect one payment does not resolve how much of an article, dataset or derived answer the publisher has licensed.

On the supply side, AWS requires a Marketplace listing with the Agentic Commerce access model and a compatible HTTPS endpoint that returns a signed x402 quote. The seller registers a verification key for the product and verifies the AWS-issued payment token before serving the response. Marketplace handles metering and pays the seller through its standard disbursement process. The seller guide says an existing registration-based product and a per-request product require separate listings. AWS Marketplace seller requirements

A publisher should describe what one quoted request delivers. Specify the freshness of a record, the fields available, what an empty result looks like and who handles a customer dispute. A price of one cent for a lookup may be perfectly clear to a machine while leaving a customer unsure whether an empty response completed the purchase they intended. A precise product description helps both a buyer comparing offers and a human reviewing an invoice.

08 — Deployment choiceStart with one buying decision

A good first task gives a purchaser and an operator the same definition of success. For example, authorize the research assistant to buy a current public-source lookup when the internal catalog has no recent entry. Select the approved product, identify the account that will pay, set a price per request and cap the session. Decide what the user will see if the price rises, the answer arrives empty or the task reaches its budget before it finds supporting evidence.

Before expanding the eligible list, review a handful of complete task histories. Look for the quoted amount, the payment identifier, delivery of the result and the evidence that justified the final answer. Include a refused quote and a session that ran out of money. A human should be able to explain those outcomes without asking the model what it meant to do.

Per-request buying makes particular sense when a paid lookup is valuable occasionally and there is little reason to maintain a separate account for each seller. A high-volume workflow with predictable demand may already have a contract or a licensed feed worth comparing at its effective cost per usable answer. Both routes can fit a well-run service. The comparison starts with how many purchases a real task needs and what the buyer receives. Comparing agent pricing units

Bring the product owner and the person who signs off the AWS bill into that review. They should be able to read the same task record and agree that the result justified the spend. Digital Applied’s AI transformation services

Operating rule

Approve the seller, the price and the task

AWS has made one paid search a practical unit of purchase on an existing cloud bill. Give an agent a session that reflects the value of the task and a product list appropriate for the information it will send. Keep the record that shows what the purchase contributed. That makes the next decision about raising the budget an ordinary product and finance discussion.

Agent workflows

Design spending rules for a paid tool

Digital Applied can help connect tool permissions, per-task costs and transaction records to the work an agent delivers.

Seller approvalTask budgetsPurchase records
Pilot inputs

Choose a task with a verifiable result

  • →A paid source the team already understands
  • →A maximum price for an individual answer
  • →A person responsible for reviewing charges
Answers

Frequently asked questions

This payment route requires a participating Marketplace product that supports x402 and a buyer payment manager configured with the AWS Billing connector. Seller restrictions and an organization’s Private Marketplace policies still apply. Ordinary Marketplace listings do not automatically accept per-request payment.

Digital Applied newsletter

Deep dives on AI, marketing and development.

Practical guides and fresh insights by email. No recycled takes.

Related dispatches

Related reading

Google Search

See more Digital Applied analysis in your Google results by adding us as a preferred source.

Add as a preferred source