When AI Agents Pay and Sellers Monetize: Building x402 Apps on AWS — Anil Nadiminti, AWS

A paywall built for humans, and traffic that is not

00:12Hello all, welcome to uh the agent e-commerce track and uh I'm Anil Lminti. I'm a senior solutions architect here at AWS. Uh I'm I'm here to talk to you today about how AWS is innovating and uh how you can build apps on uh AWS to support the agent e-commerce.

00:25So uh welcome to the session. Uh just to get you started, let me set the stage with something that you're very familiar with. Uh imagine that you are your organization is building uh a news portal like this, right? So you're all familiar with something where you're accessing the news content and then suddenly you hit a payw wall, right?

00:43So this is where uh you pull out your wallet or you try to uh figure out how to make the payments, set up your uh credentials, access keys in the sense that you you make a credit card transaction weekly, monthly or annual subscription and then get started to access the content.

00:59Right? So this is all the content that is behind a payw

Bot traffic passes human traffic

01:08wall. But uh what we see now is that uh much of the traffic that is actually being uh sent to these uh uh portals now on the internet is all coming from bots. We see that we at a infliction point where uh the bot traffic is more than the human traffic right.

01:18So it's just uh actually in fact surpassed that and uh 95% of that bot traffic is coming from AI agents. So uh essentially we are also looking at the rise of autonomous agents right so where we all started using LLMs asking questions asking for summarization being able to uh get help with using them as co-pilots getting them to do agentic work to uh do multi-step tasks and now we're in the phase of autonomous agents where agents are using the reasoning powers of large language models to complete a task and completing a task means that uh it it has to go do whatever you're asking it to do.

01:58And that's kind of where we are in the uh journey. And we see that by 2027 about a billion agents will be running uh performing tasks and 60% of the enterprises will already be uh using agentic workflows. So uh what happens when agents hit these pay walls that uh we just saw?

02:15When agents hit the pay walls, they stall, they can't operate and you see those messages that hey I cannot access content. Uh right? So at that point humans get in the loop. They try to enter and put the credit card details or API keys do the transactions for the AI agents.

02:31But all of that is manual friction, right? So essentially bringing in a human in the loop. So autonom autonomous agents actually break at that point where the uh friction is now building up. So uh now sellers of the content have uh you know couple of options right.

02:47So block all the bot traffic but by blocking all the traffic they lose this AI powered discovery they miss this uh partnership

The seller's dilemma: block bots or absorb them

03:02licensing options and AI also now supports citations right so the responses so they lose all of that uh powered citations as well if they can't sell the content essentially they lose uh revenue generating options and if you allow the bots to access those uh uh the content what it means is that you know hundreds of thousands of bots or millions of bots could be hitting your uh uh infrastructure which means that the infrastructure costs will also raise and uh you need to be able to support all of that right so you also uh when you allow bots to access content you

03:33lose attribution the IP itself right so because content is now freely available so uh both these decisions are probably not not a good option they're not ideal so there should be another ideal option where uh you would want to have your AI agents being able to get and pay for the content that they are looking for and monetize on So uh now we look at the next phase of uh uh rise in autonomous agents where agents should be able to transact and make uh discover other agents resources and essentially make payments.

04:00Right? So this is the definition of agent e-commerce where AI agents can essentially discover uh you know it's a form of e-commerce where autonomous agents can uh discover independently and make those uh uh settlements and then access content.

04:16So let's look at uh the agent e-commerce the two sides of agent e-commerce the buy side and the sell side. So uh when we talk about the buy side the agents are making these transactions and on the sell side the sellers of the content are trying to monetize on the content.

04:33So on the buy side when you look at things AI agents want to access these premium paywalled content licensed content they want to be able to hold wallets which they do not have the option today and they want to be able to make these microtransactions you just heard in the prior talk as well.

04:48But enterprises when they come to this point they want more guardrails and they do not want agents to go on spending spree. Think about it right? Would you allow your AI agents to get handled on your wallets or credit cards to be able to do that transactions and where they could go rogue as well right?

05:03So that's what the buyer side is looking at. And on the seller side uh the there are again billions of transactions that be happening with these AI bots. So sellers really want to be able to understand what kinds of bots are operating uh what kinds of transactions they're making and uh really uh do this at the edge.

05:20The sellers don't want to change their entire infrastructure and origins where the content is sitting. They want to be able to do this at the edge without changing much of this. Right? So there is again uh one common thing here on the buyer side and the seller side which is a standardized approach or a protocol to be able to solve for this uh machine to machine payments at the edge.

05:36So uh bottom line buyers are saying that they want their agents to be able to pay for content uh and uh not have humans approving this and then the sellers are saying that they want to be able to earn from the AI traffic. So bottom line the subscription model is going to change with humans in the loop to becoming humans on the loop or out of the loop and that's kind of what we are building towards.

06:01uh the the traditional one-sizefits model does not work anymore because of the fact that uh again we look at that in the next slide where uh the transactions cost will not really work right all of this needs to happening at realtime speed and the paper use and paper execution is what the f future is going to look like so if you're a seller you would have come across this right so there is a 25 cent minimum transaction fees as well as 2.5% on top of that and all of these microtransactions are uh you know in the in in like a cent subsend or you know micro cents is what we are calling them.

Why a 25 cent floor kills a microcent payment

06:38So if you add like a 25 cents on top of that it's essentially like 250 times to what you know they are essentially paying for. So the all of this model does not work and uh while we are trying to solve for that a very brief history uh of this is every HTTP call essentially responds back uh you know there's a response for that you've seen 200 status codes 404 uh you know and the 301 these are all like status codes that you're familiar with and then there is one status code which is 402 which has uh not been used it was reserved for payment required and now finally uh

07:10Coinbase has introduced this uh as uh transactions over 402 which is also called as X42 where they uh you know the protocol talks about how you can do machine to-achine transactions uh using this protocol right so we'll take a closer look at that but uh what happens within the protocol is uh you know if you look at this uh flowchart here a client makes a request to the server and then the server responds back with the

The x402 flow, end to end

07:33payment required uh the client then figures out what is the payment method that it wants to operate and then it sends the payment authorization to the server the server then utilizes a facilitator to complete the verification and then also utilizes the same facilitator to complete the transaction and once the settlement is completed onchain essentially the server will then respond back with the content right so this is what's happening under the X42 protocol I thought I'll pick one of the

07:59protocols and just uh explain this to you but uh why this is compelling is uh you know essentially there is no protocol fees uh or the fees that a consumer is paying for uh you know these microcent transactions and the merchant is paying very nominal gas fees is uh again there is zero wait time this is happening at the speed of internet and uh there is no friction there is no API keys to set up no subscriptions and the payment is the essentially the uh

08:25credential to be able to get the content so there is no centralization it's uh x42 can be extended as well and you can implement it and there are no restrictions as well so some key milestones here are you know it was introduced last year uh May 2025 it explored is not part of the Linux Foundation under open governance and it's backed by Coinbase, AWS, Google, Stripe, Anthropic, Cloudflare and Circle right so many more folks in there that are supporting that organizations in there.

08:49So uh from Amazon we have also released agent core payments uh under the bedrock uh suite so where uh agents will make be able to make payments and we'll go into some of the details here. So let's talk about the buyer side here and what is involved right?

09:06So we uh we understood from the developers that they really want to be able to get this uh agents to have wallet support. They want to be able to have real-time settlement uh have the budget and guardrails which enterprises really want and observability throughout the stack where they would want to have uh the full stack trace of everything

AgentCore Payments, and setting spend limits

09:29that's happening uh under the hood. So I'm excited to share with you that we've launched agent core payments and this is a service that allows AI agents to autonomously discover uh authorize and execute payments with a few lines of code. Uh now we've launched this in partnership with Coinbase and Stripe where you can bring wallets from Coinbase and Stripe preview to be able to do these operations and we'll go into some of the details but uh the core capabilities to start with are wallet support where you can bring the the wallets from Coinbase and Stripe.

09:55you're able to orchestrate the payments using payment connectors and uh today we support X42 with many more protocols to uh you know that are in the pipeline. The service is designed to be protocol agnostic. So as new protocols emerge, we are going to be adding the support for those protocols as well.

10:09And uh you know the the settlement is going to be instantaneous uh instant essentially and uh the payment limits can be set uh which is the most important thing that we spoke about where enterprises are looking to put some payment limits and guards on how these transactions can operate.

10:26So uh observability is builtin and uh essentially all of this uh operates with uh you know security as the uh layer that is operating uh the whole model right. So with that let's look at uh some of these uh details on how the payments limit can be set up right.

10:43So you can create payment sessions where you can set the programmatically set the maximum amount of uh uh value that can be used for transactions or you can also set expiry time in minutes. think where uh you are able to set that uh I can the agent can actually spend maybe $5 in 30 days or 60 days right so that's kind of the operation uh model that you can set with many more uh you know details that are available I'm only going over a few features but uh let's look at what happens on the buyer side when the user is asking an agent to make a particular

11:23uh you know requesting for resources right so agent completes the request by accessing tools MPPP servers other resources as well. So at that point of time if the agent uh is uh looking at you know it it finds that the there is a response from one of the tool calls or requests with a 402 agent core payments is going to handle the request uh to complete the transaction and then let the AI agent know that uh essentially the settlement happened and the AI uh agent will be able to respond back with the users.

11:54So in this process when the wallets are uh wallet support is imported the the the secret keys that you use to import the wallets actually are stored in a secure token wallet that is uh secured by KMS where you know that's there. So essentially the agent does not have access to the private keys.

12:10This is most important uh to note.

Why the agent never holds the private keys

12:16And uh next thing is that uh agent core payments is also integrated uh through uh a gateway which is also part of uh which is another service that we have to mp5 your internal APIs. Uh through agent core gateway uh the agent core payments can get access to discovery service uh in coinbase where there are 10,000 plus endpoints that are available to transact and then u again there is a per session budget that we just discussed as well.

12:43So uh there is a decoupling of agent infrastructure and uh the payment infrastructure by design where the agent can operate it it in its own loop and whenever it sees the payment the payment uh connectors orchestration payment limits and integration with third party wallets can happen right so it's important to decouple them because again skills can be poisoned inputs for the agents can also be uh you know poisoned

Decoupling payments from a poisonable agent loop

13:07by inputs as well right so where malicious actors could uh try to do that so By decoupling and making this uh by design a agents can essentially have a secure path for these transactions and payments do not touch the you know undeterministic path but this is more on a deterministic uh uh layer as well.

13:24So why this is important is that uh again agents uh the code of the agents does not have to change. You can bring your own model frameworks and then the payment itself uh can flow through uh in the payment uh uh layer itself. So again the controls the policies pending controls can be outside of the payment stack itself and again it's this is built to be protocol agnostic.

13:48So this is uh one of the console screens where it shows how you can import uh the payment connector uh and it shows that you know you can select the the coinbase wallet and the stripe preview wallet from the console and uh this is a demo in action where we are showing how a secure resource can be accessed.

14:03Now in this case uh the AI agent is essentially making a you know discovering that there is a secure source the agent core payments is kicking in and then it's completing the transaction by utilizing the wallet that is already integrated and uh the transaction completes.

14:25Now uh this is on the buyer side. Now let's look at the seller side to understand what's happening. Right? So uh again there is a lot of bot activity that's happening. We have released uh under uh the uh AWS web application firewall a feature where we have bot detection in place.

14:42Today we detect over 650 different types of bots. Think of bots like perplexity bot, GPD bot, cloud bot, you know, again Google bots, right? So there are so many bots that are out there. So we're able to detect also understand the intent of these bots.

14:51So why are these bots accessing the

Bot detection, and reading intent

15:00content? Are they accessing the content to train their models? are they doing it because they have to respond back to an intent where they're uh responding for a rag search. So we're able to identify the intent. We're also able to verify the bots and identify them by a signature.

15:08So we are able to say hey this is a verified bot. So maybe you have uh built a relation with one of these organizations and these verification will allow you to have a different pricing for the organizations that are already verified. So we'll look at that in a second.

15:26So there's also real-time traffic analysis that allows uh more uh to be customized. And I'm also happy to share with you today that we announced VAF AI traffic monetization. This is a service that allows you to monetize uh based on the content that uh you know based on how you can measure, verify and monetize based on the AI traffic that is hitting

Monetizing AI traffic at the edge

15:51your endpoints. Now if you might be familiar with CloudFront which is our content distribution network you can add a web application firewall at that point and essentially you can moni start monetizing uh right away and based on a few clicks uh you can do that again using infrastructure as code as well.

16:07Now I also spoke about a gateway service that allows you to expose your AI endpoints uh that are internal use and mcpify them. So the same web application firewalls can be used there. So your internal APIs can be MCPI and then you can start monetizing as well.

16:15So what happens during monetization? The AI agent AI bot essentially requests for some content. The bot context uh understands what kinds of bots is detecting it. It's able to detect the bot. It's able to categorize and understand the intent of the bot as we discussed earlier and verify uh and check what kind of bot is available.

16:39So then we are able to monetize uh using the X42 and the publishers get paid as well. So important to note is that again there is no SDK change no changes at the origin origin. Publishers keep 100% of the revenue as well and uh again there is no transaction fees or subscription fees.

17:00So this supports X42 and we are adding support for more uh protocols as well. U a few dimensions on how you can start monetizing. Think uh you have separate paths. So a slash uh uh blog in this case can be charging for a different rate than a slash research or maybe an API endpoint itself and uh you you know the identity of these bots.

17:24Again, if you make uh some kind of a relationship with the uh bots uh companies, organizations, maybe you make a relation with Anthropic, then you can essentially have a different pricing for those bots versus different uh unverified bots, right?

17:30So, uh think of that option. And then you can also set different pricing for intent as well. If somebody's coming here, if a bot is accessing the content for uh again

Pricing by path, by identity, by intent

17:47training, you can charge a different rate than what it's doing for a search as well. So again, these are different VAF rules. they can be uh in a combination of end or or then you can access that. So this is how the reimagine flow would look like uh where you're allowing the AI agents or you know essentially verified bots and unverified bots to have different pricing and humans to have different pricing.

18:09Some cases you want to have humans to access the content freely. Some cases again the humans could be charged where the bots could be charged differently as well. Right? So this is how uh you know you can reimagine the price. So again there is some uh dashboards that show the revenue numbers and how uh you know you're uh able to aggregate by different uh bots and figure out what kind of revenue model you want to operate and it also shows what is the path uh that's being accessed by these bots right so uh again

What the volume looks like today

18:38what is currently uh everyone using agent e-commerce for they're using agent commerce to uh run um again LLM inference getting uh compute web scraping uh they're uh creating research search agents to be able to uh you know serve the requests and uh agent to agent as well.

18:53We see MCPs also being monetized now. Uh again this is the last 12 months of uh traffic uh from um again what we are seeing on uh Coinbase agentic market u where you're seeing that a $50 million volume transaction happened over 170 million transactions.

19:13The average settlement time is 200 milliseconds on base uh with about a tenth of a cent as cost per transaction. So I spoke to you about agent uh agent core payments which is uh one of the you know parts of the bigger ecosystem agent uh bed agent core uh where you can essentially bring your own model.

19:30You can bring your own framework and uh you can start building AI agents. You can add context by adding memory. You can uh bring again your own managed knowledge bases. You can add web search capabilities to the agents. you can mcpify your internal APIs and then you can uh have many more features like being able to run evaluation on how your agents are performing.

19:53So again you can uh use runtime which is a bedrock uh agent core runtime where you can bring your own agentic uh uh application and serve uh at scale and every request will have uh its own isolated uh micro virtual machine that is running to serve the requests.

20:10So that's it from my side here today. Uh thank you and uh hope you have a nice day.