{
  "video_id": "qTZirYu9pr0",
  "channel_slug": "aidotengineer",
  "channel_handle": "aidotengineer",
  "title": "When AI Agents Pay and Sellers Monetize: Building x402 Apps on AWS — Anil Nadiminti, AWS",
  "duration_seconds": 1241.0,
  "url": "https://www.youtube.com/watch?v=qTZirYu9pr0",
  "upload_date": "",
  "transcript": "[music]\nHello all, welcome to uh the agent\ne-commerce track and uh I'm Anil Lminti.\nI'm a senior solutions architect here at\nAWS. Uh I'm I'm here to talk to you\ntoday about how AWS is innovating and uh\nhow you can build apps on uh AWS to\nsupport the agent e-commerce. So uh\nwelcome to the session. Uh just to get\nyou started, let me set the stage with\nsomething that you're very familiar\nwith. Uh imagine that you are your\norganization is building uh a news\nportal like this, right? So you're all\nfamiliar with something where you're\naccessing the news content and then\nsuddenly you hit a payw wall, right? So\nthis is where uh you pull out your\nwallet or you try to uh figure out how\nto make the payments, set up your uh\ncredentials, access keys in the sense\nthat you you make a credit card\ntransaction weekly, monthly or annual\nsubscription and then get started to\naccess the content. Right? So this is\nall the content that is behind a payw\nwall. But uh what we see now is that uh\nmuch of the traffic that is actually\nbeing uh sent to these uh uh portals now\non the internet is all coming from bots.\nWe see that we at a infliction point\nwhere uh the bot traffic is more than\nthe human traffic right. So it's just uh\nactually in fact surpassed that and uh\n95% of that bot traffic is coming from\nAI agents. So uh essentially we are also\nlooking at the rise of autonomous agents\nright so where we all started using LLMs\nasking questions asking for\nsummarization being able to uh get help\nwith using them as co-pilots getting\nthem to do agentic work to uh do\nmulti-step tasks and now we're in the\nphase of autonomous agents where agents\nare using the reasoning powers of large\nlanguage models to complete a task and\ncompleting a task means that uh it it\nhas to go do whatever you're asking it\nto do. And that's kind of where we are\nin the uh journey. And we see that by\n2027 about a billion agents will be\nrunning uh performing tasks and 60% of\nthe enterprises will already be uh using\nagentic workflows. So uh what happens\nwhen agents hit these pay walls that uh\nwe just saw? When agents hit the pay\nwalls, they stall, they can't operate\nand you see those messages that hey I\ncannot access content. Uh right? So at\nthat point humans get in the loop. They\ntry to enter and put the credit card\ndetails or API keys do the transactions\nfor the AI agents. But all of that is\nmanual friction, right? So essentially\nbringing in a human in the loop. So\nautonom autonomous agents actually break\n[clears throat] at that point where the\nuh friction is now building up. So uh\nnow sellers of the content have uh you\nknow couple of options right. So block\nall the bot traffic but by blocking all\nthe traffic they lose this AI powered\ndiscovery they miss this uh partnership\nlicensing options and AI also now\nsupports citations right so the\nresponses so they lose all of that uh\npowered citations as well if they can't\nsell the content essentially they lose\nuh revenue generating options and if you\nallow the bots to access those uh uh the\ncontent what it means is that you know\nhundreds of thousands of bots or\nmillions of bots could be hitting your\nuh uh infrastructure which means that\nthe infrastructure costs will also raise\nand uh you need to be able to support\nall of that right so you also uh when\nyou allow bots to access content you\nlose attribution the IP itself right so\nbecause content is now freely available\nso uh both these decisions are probably\nnot not a good option they're not ideal\nso there should be another ideal option\nwhere uh you would want to have your AI\nagents being able to get and pay for the\ncontent that they are looking for and\nmonetize on\nSo uh now we look at the next phase of\nuh uh rise in autonomous agents where\nagents should be able to transact and\nmake uh discover other agents resources\nand essentially make payments. Right? So\nthis is the definition of agent\ne-commerce where AI agents can\nessentially discover uh you know it's a\nform of e-commerce where autonomous\nagents can uh discover independently and\nmake those uh uh settlements and then\naccess content. So let's look at uh the\nagent e-commerce the two sides of agent\ne-commerce the buy side and the sell\nside. So uh when we talk about the buy\nside the agents are making these\ntransactions and on the sell side the\nsellers of the content are trying to\nmonetize on the content. So on the buy\nside when you look at things AI agents\nwant to access these premium paywalled\ncontent licensed content they want to be\nable to hold wallets which they do not\nhave the option today and they want to\nbe able to make these microtransactions\nyou just heard in the prior talk as\nwell. But enterprises when they come to\nthis point they want more guardrails and\nthey do not want agents to go on\nspending spree. Think about it right?\nWould you allow your AI agents to get\nhandled on your wallets or credit cards\nto be able to do that transactions and\nwhere they could go rogue as well right?\nSo that's what the buyer side is looking\nat. And on the seller side uh the there\nare again billions of transactions that\nbe happening with these AI bots. So\nsellers really want to be able to\nunderstand what kinds of bots are\noperating uh what kinds of transactions\nthey're making and uh really uh do this\nat the edge. The sellers don't want to\nchange their entire infrastructure and\norigins where the content is sitting.\nThey want to be able to do this at the\nedge without changing much of this.\nRight? So there is again uh one common\nthing here on the buyer side and the\nseller side which is a standardized\napproach or a protocol to be able to\nsolve for this uh machine to machine\npayments at the edge. So uh bottom line\nbuyers are saying that they want their\nagents to be able to pay for content uh\nand uh not have humans approving this\nand then the sellers are saying that\nthey want to be able to earn from the AI\ntraffic. So bottom line the subscription\nmodel is going to change with humans in\nthe loop to becoming humans on the loop\nor out of the loop and that's kind of\nwhat we are building towards. uh the the\ntraditional one-sizefits model does not\nwork anymore because of the fact that uh\nagain we look at that in the next slide\nwhere uh the transactions cost will not\nreally work right all of this needs to\nhappening at realtime speed and the\npaper use and paper execution is what\nthe f future is going to look like so if\nyou're a seller you would have come\nacross this right so there is a 25 cent\nminimum transaction fees as well as 2.5%\non top of that and all of these\nmicrotransactions are uh you know in the\nin in like a cent subsend or you know\nmicro cents is what we are calling them.\nSo if you add like a 25 cents on top of\nthat it's essentially like 250 times to\nwhat you know they are essentially\npaying for. So the all of this model\ndoes not work and uh while we are trying\nto solve for that a very brief history\nuh of this is every HTTP call\nessentially responds back uh you know\nthere's a response for that you've seen\n200 status codes 404 uh you know and the\n301 these are all like status codes that\nyou're familiar with and then there is\none status code which is 402 which has\nuh not been used it was reserved for\npayment required and now finally uh\nCoinbase has introduced this uh as uh\ntransactions over 402 which is also\ncalled as X42 where they uh you know the\nprotocol talks about how you can do\nmachine to-achine transactions uh using\nthis protocol right so we'll take a\ncloser look at that but uh what happens\nwithin the protocol is uh you know if\nyou look at this uh flowchart here a\nclient makes a request to the server and\nthen the server responds back with the\npayment required uh the client then\nfigures out what is the payment method\nthat it wants to operate and then it\nsends the payment authorization to the\nserver the server then utilizes a\nfacilitator to complete the verification\nand then also utilizes the same\nfacilitator to complete the transaction\nand once the settlement is completed\nonchain essentially the server will then\nrespond back with the content right so\nthis is what's happening under the X42\nprotocol I thought I'll pick one of the\nprotocols and just uh explain this to\nyou but uh why this is compelling is uh\nyou know essentially there is no\nprotocol fees uh or the fees that a\nconsumer is paying for uh you know these\nmicrocent transactions and the merchant\nis paying very nominal gas fees is uh\nagain there is zero wait time this is\nhappening at the speed of internet and\nuh there is no friction there is no API\nkeys to set up no subscriptions and the\npayment is the essentially the uh\ncredential to be able to get the content\nso there is no centralization it's uh\nx42 can be extended as well and you can\nimplement it and there are no\nrestrictions as well so some key\nmilestones here are you know it was\nintroduced last year uh May 2025 it\nexplored is not part of the Linux\nFoundation under open governance and\nit's backed by Coinbase, AWS,\nGoogle, Stripe, Anthropic, Cloudflare\nand Circle right so many more folks in\nthere that are supporting that\norganizations in there. So uh from\nAmazon we have also released agent core\npayments uh under the bedrock uh suite\nso where uh agents will make be able to\nmake payments and we'll go into some of\nthe details here. So let's talk about\nthe buyer side here and what is involved\nright? So we uh we understood from the\ndevelopers that they really want to be\nable to get this uh agents to have\nwallet support. They want to be able to\nhave real-time settlement uh have the\nbudget and guardrails which enterprises\nreally want and observability throughout\nthe stack where they would want to have\nuh the full stack trace of everything\nthat's happening uh under the hood. So\nI'm excited to share with you that we've\nlaunched agent core payments and this is\na service that allows AI agents to\nautonomously discover uh authorize and\nexecute payments with a few lines of\ncode. Uh now we've launched this in\npartnership with Coinbase and Stripe\nwhere you can bring wallets from\nCoinbase and Stripe preview to be able\nto do these operations and we'll go into\nsome of the details but uh the core\ncapabilities to start with are wallet\nsupport where you can bring the the\nwallets from Coinbase and Stripe. you're\nable to orchestrate the payments using\npayment connectors and uh today we\nsupport X42 with many more protocols to\nuh you know that are in the pipeline.\nThe service is designed to be protocol\nagnostic. So as new protocols emerge, we\nare going to be adding the support for\nthose protocols as well. And uh you know\nthe the settlement is going to be\ninstantaneous uh instant essentially and\nuh the payment limits can be set uh\nwhich is the most important thing that\nwe spoke about where enterprises are\nlooking to put some payment limits and\nguards on how these transactions can\noperate. So uh observability is builtin\nand uh essentially all of this uh\noperates with uh you know security as\nthe uh layer that is operating uh the\nwhole model right. So with that let's\nlook at uh some of these uh details on\nhow the payments limit can be set up\nright. So you can create payment\nsessions where you can set the\nprogrammatically set the maximum amount\nof uh uh value that can be used for\ntransactions or you can also set expiry\ntime in minutes. think where uh you are\nable to set that uh I can the agent can\nactually spend maybe $5 in 30 days or 60\ndays right so that's kind of the\noperation uh model that you can set with\nmany more uh you know details that are\navailable I'm only going over a few\nfeatures but uh let's look at what\nhappens on the buyer side when the user\nis asking an agent to make a particular\nuh you know requesting for resources\nright so agent completes the request by\naccessing tools MPPP servers other\nresources as well. So at that point of\ntime if the agent uh is uh looking at\nyou know it it finds that the there is a\nresponse from one of the tool calls or\nrequests with a 402 agent core payments\nis going to handle the request uh to\ncomplete the transaction and then let\nthe AI agent know that uh essentially\nthe settlement happened and the AI uh\nagent will be able to respond back with\nthe users. So in this process when the\nwallets are uh wallet support is\nimported the the the secret keys that\nyou use to import the wallets actually\nare stored in a secure token wallet that\nis uh secured by KMS where you know\nthat's there. So essentially the agent\ndoes not have access to the private\nkeys. This is most important uh to note.\nAnd uh next thing is that uh agent core\npayments is also integrated uh through\nuh a gateway which is also part of uh\nwhich is another service that we have to\nmp5 your internal APIs. Uh through agent\ncore gateway uh the agent core payments\ncan get access to discovery service uh\nin coinbase where there are 10,000 plus\nendpoints that are available to transact\nand then u again there is a per session\nbudget that we just discussed as well.\nSo uh there is a decoupling of agent\ninfrastructure and uh the payment\ninfrastructure by design where the agent\ncan operate it it in its own loop and\nwhenever it sees the payment the payment\nuh connectors orchestration payment\nlimits and integration with third party\nwallets can happen right so it's\nimportant to decouple them because again\nskills can be poisoned inputs for the\nagents can also be uh you know poisoned\nby inputs as well right so where\nmalicious actors could uh try to do that\nso By decoupling and making this uh by\ndesign a agents can essentially have a\nsecure path for these transactions and\npayments do not touch the you know\nundeterministic path but this is more on\na deterministic uh uh layer as well. So\nwhy this is important is that uh again\nagents uh the code of the agents does\nnot have to change. You can bring your\nown model frameworks and then the\npayment itself uh can flow through uh in\nthe payment uh uh layer itself. So again\nthe controls the policies pending\ncontrols can be outside of the payment\nstack itself and again it's this is\nbuilt to be protocol agnostic. So this\nis uh one of the console screens where\nit shows how you can import uh the\npayment connector uh and it shows that\nyou know you can select the the coinbase\nwallet and the stripe preview wallet\nfrom the console and uh this is a demo\nin action where we are showing how a\nsecure resource can be accessed. Now in\nthis case uh the AI agent is essentially\nmaking a you know discovering that there\nis a secure source the agent core\npayments is kicking in and then it's\ncompleting the transaction by utilizing\nthe wallet that is already integrated\nand uh the transaction completes. Now uh\nthis is on the buyer side. Now let's\nlook at the seller side to understand\nwhat's happening. Right? So uh again\nthere is a lot of bot activity that's\nhappening. We have released uh under uh\nthe uh AWS web application firewall a\nfeature where we have bot detection in\nplace. Today we detect over 650\ndifferent types of bots. Think of bots\nlike perplexity bot, GPD bot, cloud bot,\nyou know, again Google bots, right? So\nthere are so many bots that are out\nthere. So we're able to detect also\nunderstand the intent of these bots. So\nwhy are these bots accessing the\ncontent? Are they accessing the content\nto train their models? are they doing it\nbecause they have to respond back to an\nintent where they're uh responding for a\nrag search. So we're able to identify\nthe intent. We're also able to verify\nthe bots and identify them by a\nsignature. So we are able to say hey\nthis is a verified bot. So maybe you\nhave uh built a relation with one of\nthese organizations and these\nverification will allow you to have a\ndifferent pricing for the organizations\nthat are already verified. So we'll look\nat that in a second. So there's also\nreal-time traffic analysis that allows\nuh more uh to be customized. And I'm\nalso happy to share with you today that\nwe announced VAF AI traffic\nmonetization. This is a service that\nallows you to monetize uh based on the\ncontent that uh you know based on how\nyou can measure, verify and monetize\nbased on the AI traffic that is hitting\nyour endpoints. Now if you might be\nfamiliar with CloudFront which is our\ncontent distribution network you can add\na web application firewall at that point\nand essentially you can moni start\nmonetizing uh right away and based on a\nfew clicks uh you can do that again\nusing infrastructure as code as well.\nNow I also spoke about a gateway service\nthat allows you to expose your AI\nendpoints uh that are internal use and\nmcpify them. So the same web application\nfirewalls can be used there. So your\ninternal APIs can be MCPI and then you\ncan start monetizing as well. So what\nhappens during monetization? The AI\nagent AI bot essentially requests for\nsome content. The bot context uh\nunderstands what kinds of bots is\ndetecting it. It's able to detect the\nbot. It's able to categorize and\nunderstand the intent of the bot as we\ndiscussed earlier and verify uh and\ncheck what kind of bot is available. So\nthen we are able to monetize uh using\nthe X42 and the publishers get paid as\nwell. So important to note is that again\nthere is no SDK change no changes at the\norigin origin. Publishers keep 100% of\nthe revenue as well and uh again there\nis no transaction fees or subscription\nfees. So this supports X42 and we are\nadding support for more uh protocols as\nwell. U a few dimensions on how you can\nstart monetizing. Think uh you have\nseparate paths. So a slash uh uh blog in\nthis case can be charging for a\ndifferent rate than a slash research or\nmaybe an API endpoint itself and uh you\nyou know the identity of these bots.\nAgain, if you make uh some kind of a\nrelationship with the uh bots uh\ncompanies, organizations, maybe you make\na relation with Anthropic, then you can\nessentially have a different pricing for\nthose bots versus different uh\nunverified bots, right? So, uh think of\nthat option. And then you can also set\ndifferent pricing for intent as well. If\nsomebody's coming here, if a bot is\naccessing the content for uh again\ntraining, you can charge a different\nrate than what it's doing for a search\nas well. So again, these are different\nVAF rules. they can be uh in a\ncombination of end or or then you can\naccess that. So this is how the\nreimagine flow would look like uh where\nyou're allowing the AI agents or you\nknow essentially verified bots and\nunverified bots to have different\npricing and humans to have different\npricing. Some cases you want to have\nhumans to access the content freely.\nSome cases again the humans could be\ncharged where the bots could be charged\ndifferently as well. Right? So this is\nhow uh you know you can reimagine the\nprice. So again there is some uh\ndashboards that show the revenue numbers\nand how uh you know you're uh able to\naggregate by different uh bots and\nfigure out what kind of revenue model\nyou want to operate and it also shows\nwhat is the path uh that's being\naccessed by these bots right so uh again\nwhat is currently uh everyone using\nagent e-commerce for they're using agent\ncommerce to uh run um again LLM\ninference getting uh compute web\nscraping uh they're uh creating research\nsearch agents to be able to uh you know\nserve the requests and uh agent to agent\nas well. We see MCPs also being\nmonetized now. Uh again this is the last\n12 months of uh traffic uh from um again\nwhat we are seeing on uh Coinbase\nagentic market u where you're seeing\nthat a $50 million volume transaction\nhappened over 170 million transactions.\nThe average settlement time is 200\nmilliseconds on base uh with about a\ntenth of a cent as cost per transaction.\nSo I spoke to you about agent uh agent\ncore payments which is uh one of the you\nknow parts of the bigger ecosystem agent\nuh bed agent core uh where you can\nessentially bring your own model. You\ncan bring your own framework and uh you\ncan start building AI agents. You can\nadd context by adding memory. You can uh\nbring again your own managed knowledge\nbases. You can add web search\ncapabilities to the agents. you can\nmcpify your internal APIs and then you\ncan uh have many more features like\nbeing able to run evaluation on how your\nagents are performing. So again you can\nuh use runtime which is a bedrock uh\nagent core runtime where you can bring\nyour own agentic uh uh application and\nserve uh at scale and every request will\nhave uh its own isolated uh micro\nvirtual machine that is running to serve\nthe requests. So that's it from my side\nhere today. Uh thank you and uh hope you\nhave a nice day.\n[applause]\n>> [music]",
  "transcript_chars": 20516,
  "ingested_at": "2026-09-03T10:30:15.808218+00:00",
  "source": "channel",
  "yt_meta": {
    "view_count": null,
    "like_count": null,
    "channel_id": null,
    "categories": null,
    "tags": null
  }
}