Introducing Visa Clever Commerce on AWS: Enabling agentic commerce with Amazon Bedrock AgentCore

This put up is cowritten with Sangeetha Bharath and Seemal Zaman from Visa.

Throughout each business, agentic AI is redefining how work will get completed by shifting digital experiences from guide, user-driven interactions to autonomous, outcome-driven workflows. In contrast to conventional AI methods that merely reply questions or present solutions, agentic AI introduces clever brokers able to reasoning, appearing, collaborating with different brokers, and finishing multistep duties on the consumer’s behalf. This shift is already reworking sectors akin to journey, healthcare, banking, logistics, and customer support, the place brokers can analysis, plan, optimize, and execute end-to-end processes with minimal human intervention.

The funds business can be coming into a serious transformation as agentic commerce shifts how shoppers and companies provoke, authorize, and full transactions. As an alternative of guide steps throughout a number of apps and web sites, autonomous brokers can coordinate discovery, decision-making, and safe funds within the background. This alteration mirrors the shift to ecommerce within the early 2000s, when digital checkout reshaped buyer expectations. At this time, agentic commerce is establishing an analogous basis by making funds extra seamless, contextual, and clever. Builders constructing brokers require enablement and help for this to essentially work securely and at scale. Till lately, even superior AI brokers might solely help with planning, evaluating choices, and getting ready carts, however to finish the acquisition, shoppers nonetheless needed to pull out their credit score or debit playing cards and finalize the checkout course of themselves. With out trusted, standardized, and compliant integrations to current fee networks, brokers couldn’t reliably or safely provoke transactions from finish to finish.

To help this shift, Amazon Net Providers (AWS) and Visa have teamed as much as assist enterprises serving each shoppers and business-to-business (B2B) enterprises construct for the brand new agentic commerce world. In April 2025, Visa launched Visa Clever Commerce for builders (and even non builders) to attach their agentic fee purposes on to Visa’s fee community and set off transactions with simple, pure language instructions. This initiative and providing additionally gives help for constructing network-agnostic agentic commerce flows and permits safe communication between brokers and retailers utilizing Visa’s Trusted Agent Protocol.

As a part of this partnership with AWS, Visa is planning to make use of Amazon Bedrock AgentCore to host Mannequin Context Protocol (MCP) instruments so companions can construct end-to-end, community agnostic agentic workflows on the platform.

The objective is easy: present a safe, scalable basis for constructing the following technology of clever commerce options.

Introducing Visa Clever Commerce on AWS

Visa Clever Commerce empowers companies and builders to construct the following technology of agentic fee experiences. In AWS Market, builders can find out about Visa’s suite of important agentic commerce instruments, together with authentication, agentic tokenization, and consumer intent seize, which can be designed to allow autonomous, safe, and contextual fee flows. The blueprints Visa and AWS are making accessible by the Amazon Bedrock AgentCore samples repo are designed for integration with Visa’s MCP server and agentic APIs, supporting safe, tokenized transactions, with further help for multi-network agentic commerce flows coming quickly. Via this initiative, Visa and AWS are making it simpler and extra accessible to include funds into agentic workflows.

How Amazon Bedrock AgentCore powers these options

Earlier than diving into the precise use circumstances, it’s essential to know the position Amazon Bedrock AgentCore performs because the foundational infrastructure enabling these agentic commerce experiences. Bedrock AgentCore isn’t merely one other part—it’s the safe, scalable spine that makes production-grade multi-agent methods attainable.

The worth Amazon Bedrock AgentCore provides:

• The core of this resolution is Amazon Bedrock AgentCore Runtime, a safe, serverless internet hosting atmosphere purpose-built for AI brokers and MCP servers. Every agent runs in remoted micro digital machine (VM) sandboxes so delicate information akin to journey itineraries, fee credentials, and personally identifiable info (PII) stays protected all through workflows. Bedrock AgentCore Runtime scales mechanically to deal with hundreds of concurrent customers with out guide capability planning, serving peak vacation visitors as simply as low season queries. In contrast to typical request-response APIs, it helps lengthy session durations and huge context payloads, enabling brokers to keep up context throughout multiday journey planning or advanced comparability buying classes.
• Supporting Bedrock AgentCore Runtime is Amazon Bedrock AgentCore Identification, offering inbound authentication (utilizing AWS Amplify on this resolution) for consumer sign-in and outbound authentication to securely entry endpoints akin to on-line journey company (OTA) or retail MCP servers, and Visa Clever Commerce MCP enabling the consumer’s id, consent state, and approved fee credentials to be securely carried by the workflow with out exposing delicate info.
• Via Amazon Bedrock AgentCore Gateway, brokers acquire ruled, auditable entry to instruments and MCP servers for flight and resort search, product search, and Visa Clever Commerce MCP for fee, consent, and card lifecycle operations. Amazon Bedrock AgentCore Gateway processes agent instrument requests and permits instrument calls to fulfill the trusted entry controls required in regulated fee flows.
• Amazon Bedrock AgentCore Reminiscence maintains long-duration context over prolonged, multistep journeys like journey planning, product analysis, and checkout. This allows brokers to motive extra successfully and bear in mind advanced information—akin to multicity itineraries, resort bundles, climate insights, or service provider presents—with out efficiency affect. The present implementation makes use of Bedrock AgentCore short-term reminiscence to keep up conversational context and session state whereas additionally enabling safe, managed context sharing throughout brokers. A future replace will incorporate long-term reminiscence capabilities for extracting consumer preferences and producing session summaries.
• To satisfy regulatory and compliance necessities, Amazon Bedrock AgentCore Observability, constructed on OpenTelemetry (OTEL), gives full transparency into agent operations throughout the whole workflow by capturing an entire, audit-ready file of each agent motion, together with reasoning traces, particular person spans, instrument invocations, MCP server calls, authentication flows, and latency metrics. Customers can view these operations within the Amazon CloudWatch generative AI observability dashboard.

A reusable supervisor structure

One of many key architectural benefits demonstrated in these samples is the reusable supervisor sample. This sample makes use of the brokers as instruments paradigm, the place subagents are uncovered to the supervisor as instruments it could possibly invoke. Each the journey reserving and buying assistant options share the identical supervisor agent design, which acts because the central orchestrator coordinating consumer interactions. How the shared supervisor works:

• Routes requests to the suitable specialised brokers primarily based on consumer intent
• Maintains dialog context utilizing Amazon Bedrock AgentCore Reminiscence by the Strands AgentCore Reminiscence Session Supervisor, preserving state throughout classes
• Codecs and presents responses from subagents again to customers
• Handles multiturn conversations with full context consciousness

This implies you may successfully deploy the identical supervisor infrastructure for each journey and buying use circumstances—or every other agentic commerce situation—merely by swapping out the specialised subagents (and making desired system immediate updates). The supervisor’s orchestration logic, reminiscence administration, and dialog dealing with stay constant. That is essential as a result of this modular method reduces improvement overhead. Relatively than constructing separate orchestration methods for every use case, builders can:

• Reuse the supervisor agent throughout a number of domains
• Add new specialised brokers (akin to insurance coverage, automotive rental, or grocery) with out modifying the core orchestration
• Preserve constant consumer expertise patterns throughout completely different commerce situations
• Use the identical Amazon Bedrock AgentCore infrastructure (akin to Runtime, Reminiscence, Identification, Gateway, or Observability) for his or her deployments

This put up incorporates two multi-agent samples utilizing Visa Clever Commerce:

1. Journey reserving agent
2. Buying assistant agent

Half 1: Reimagining the journey reserving expertise with Amazon Bedrock AgentCore and Visa

Vacationers face a disjointed journey planning expertise, leaping throughout airline websites, OTAs, resort platforms, loyalty portals, evaluate channels, and fee screens to plan a single journey. Costs fluctuate by the minute, loyalty program phrases are advanced, personalization is inconsistent, and even after hours of analysis, there’s no assure they discovered the best choice or maximized worth from their card advantages.

To deal with these challenges, we’ve developed a journey reserving multi-agent system utilizing Amazon Bedrock AgentCore, Strands Brokers, and Visa Clever Commerce that plans, optimizes, and books end-to-end journey experiences on a consumer’s behalf with governance and safety in place. It brings collectively discovery, personalization, and safe funds right into a single, seamless workflow pushed by pure language.

Amazon Bedrock AgentCore gives a safe, serverless runtime purpose-built for orchestrating multi-agent methods with long-running classes, giant context payloads, and ruled instrument entry. Its built-in isolation, id, observability, and MCP integration make it effectively fitted to dealing with delicate journey and fee interactions at manufacturing scale.

With Visa Clever Commerce, the consumer can approve or verify their intent, permitting the identical multi-agent system that builds the itinerary to authorize bookings and execute fee, making a seamless and extremely private journey commerce expertise that goes far past conventional journey analysis brokers.

The journey agent blueprint consists of three specialised brokers working collectively to supply complete journey planning:

1. Supervisor – Most important agentic orchestrator that coordinates interactions
2. Journey assistant – Handles journey planning, bookings, and vacation spot info
3. Cart supervisor – Manages buying cart, funds, and buy stream

The supervisor acts because the central orchestrator of the whole expertise. It orchestrates conversations, delegates duties to specialised brokers, and manages the consumer’s itinerary. Working on Amazon Bedrock AgentCore Runtime, the supervisor agent maintains dialog context throughout classes utilizing AgentCore Reminiscence, enabling it to recollect consumer preferences, in-progress itineraries, and prior choices even throughout prolonged planning classes. Core capabilities embody:

• Routes consumer requests to acceptable specialised brokers
• Maintains dialog context and reminiscence throughout classes utilizing AgentCore Reminiscence
• Codecs and presents responses from subagents to customers
• Handles multiturn conversations with context consciousness
• Helps a number of merchandise sorts akin to flight, resort, exercise, restaurant, and transport

The journey assistant focuses on travel-related queries together with vacation spot analysis, climate info, flight and resort searches, and native suggestions utilizing OTA MCP instruments by the Amazon Bedrock AgentCore Gateway. It compares presents, assembles itineraries, manages modifications, and aligns journey parts—air, resort, actions—with consumer preferences and constraints. Though these OTA instruments aren’t inherent to MCP servers, we are able to use Amazon Bedrock AgentCore Runtime to host them and expose them to the brokers as MCP appropriate instruments utilizing AgentCore Gateway.

Instruments:

• Climate info – get_weather(question)
• Web search – search_tool(question)
• Native locations search – google_places_tool(question)
• Flight search – get_flight_offers_tool(origin, vacation spot, departure_date, adults, max_price, forex)
• Lodge search – get_hotel_data_tool(city_code, rankings, facilities, max_price)
• Date updates – update_itinerary_date(user_id, identifier, item_type, new_date)

The cart supervisor handles buying cart operations, fee processing, and buy stream. That is the place Amazon Bedrock AgentCore safety capabilities turn out to be important. Bedrock AgentCore Identification manages the safe handoff to Visa Clever Commerce, enabling consumer id, consent state, and tokenized credentials to stream by the fee authorization with out exposing delicate information. Bedrock AgentCore Runtime remoted execution runs fee operations in protected sandboxes, and Bedrock AgentCore Observability captures the entire transaction stream for regulatory compliance and audit necessities.

The next diagram illustrates this structure.

Earlier than continuing with fee, the agent requests human affirmation the place the consumer units clear parameters and permits the agent to spend on their behalf. The request_purchase_confirmation instrument first captures the consumer’s express authorization, after which the confirm_purchase instrument completes the transaction when the authorization has been secured. The agent then makes use of Visa Clever Commerce APIs to request fee credentials, set off authentication, and full the acquisition securely. This human-in-the-loop step implies that customers retain management whereas benefiting from agentic automation.

Instruments:

• Cart viewing – get_cart(user_id)
• Including objects to cart – add_to_cart(user_id, objects)
• Eradicating objects – remove_from_cart(user_id, identifiers, item_type)
• Cost card administration – onboard_card(user_id, card_number, expiration_date, cvv, card_type, is_primary), request_purchase_confirmation, confirm_purchase

Going additional with Expedia Group’s Speedy APIs

To increase this pattern structure, contemplate integrating Expedia Group’s Speedy APIs to allow flight, lodging, automotive rental, and exercise bookings. These APIs ship real-time entry to world journey stock, supporting richer itineraries and seamless end-to-end reserving experiences. Speedy APIs will be built-in immediately or by utilizing MCP servers, offering flexibility and alignment together with your architectural and scalability wants.

To be taught extra, go to Expedia Group Speedy API Developer Hub.

Half 2: Way forward for buying with agentic commerce, powered by Amazon Bedrock AgentCore and Visa

With so many on-line portals, buying apps, loyalty packages, and checkout flows competing for consideration, customers should navigate a posh maze to purchase a single merchandise. If you obtain a promotional provide, the product hyperlink takes you to a unique website, loyalty factors disguise in one more portal, and checkout requires reentering the identical card particulars throughout a number of retailers. Costs shift continually, availability adjustments by the hour, and even after evaluating every thing manually, customers nonetheless really feel not sure whether or not they obtained one of the best deal, the quickest supply, or the utmost worth from their rewards.

With a multi-agent buying assistant powered by Amazon Bedrock AgentCore and built-in with Visa Clever Commerce, buying turns into frictionless because the work shifts from the patron to the brokers. As an alternative of juggling tabs and evaluating costs, customers can say one thing like: “Discover one of the best provide for Sony PlayStation 5 Professional, evaluate it throughout retailers for Black Friday promotions, verify supply dates, apply my rewards. My Price range is below $500.” Behind the scenes, a coordinated crew of brokers will search the product throughout varied service provider websites and portals, verify and evaluate pricing together with promotions, evaluate supply timelines and apply loyalty advantages.

With Visa Clever Commerce built-in, the buying assistant can validate the consumer’s id, retrieve tokenized credentials tied to their particular request, and execute the acquisition with out the consumer navigating a single checkout web page. Your complete buying stream, from analysis to comparability to optimization to fee, occurs autonomously, with the consumer guiding the method by pure language as a substitute of guide clicks. The buying assistant agent blueprint consists of three specialised brokers working collectively to supply complete buying planning:

1. Supervisor – Most important orchestrator that coordinates interactions
2. Buying – Handles product search and proposals
3. Cart supervisor – Manages buying cart, funds, and buy stream

As highlighted at first, we’re utilizing a reusable supervisor agent structure for each journey assistant and buying assistant options. For the multi-agent buying assistant use case, the supervisor agent acts because the central orchestrator of the whole expertise. It orchestrates conversations, delegates duties to specialised brokers, and manages the consumer’s itinerary. Working on Amazon Bedrock AgentCore Runtime, the supervisor agent maintains dialog context throughout classes utilizing Bedrock AgentCore Reminiscence, enabling it to recollect consumer preferences, in-progress inineraries, and prior choices even throughout prolonged planning classes.

Core capabilities embody:

• Routes consumer requests to acceptable specialised brokers
• Maintains dialog context and reminiscence throughout classes utilizing Bedrock AgentCore Reminiscence
• Codecs and presents responses from sub-agents to customers
• Handles multiturn conversations with context consciousness

The buying assistant focuses on product discovery, suggestions, and packing listing technology. Utilizing Amazon Bedrock AgentCore Gateway, it connects to retail MCP servers for product search whereas sustaining audit trails of every question. Bedrock AgentCore Reminiscence preserves buying context—remembering finances constraints, most well-liked manufacturers, and objects already thought-about—throughout the whole buying journey.

Instruments:

• Product search – single_productsearch(user_id, query)
• Packing listing technology – generate_packinglist(user_id, query)

The cart supervisor handles buying cart operations, fee processing, and buy stream. That is the place Amazon Bedrock AgentCore safety capabilities turn out to be important. Bedrock AgentCore Identification manages the safe handoff to Visa Clever Commerce, enabling consumer id, consent state, and tokenized credentials to stream by the fee authorization with out exposing delicate information. Bedrock AgentCore Runtime remoted execution runs fee operations in protected sandboxes, and Bedrock AgentCore Observability captures the transaction stream and interactions with Amazon Bedrock AgentCore Runtime, Bedrock AgentCore Reminiscence, and Bedrock AgentCore Gateway for regulatory compliance and audit necessities.

The next diagram exhibits this structure.

Earlier than continuing with fee, the agent requests human affirmation the place the consumer units clear parameters and permits the agent to spend on their behalf. The agent then makes use of Visa Clever Commerce APIs to request fee credentials, set off authentication, and full the acquisition securely. This human-in-the-loop step offers customers management whereas benefiting from agentic automation.

Instruments:

• Cart viewing – get_cart(user_id)
• Including objects to cart – add_to_cart(user_id, objects)
• Eradicating objects – remove_from_cart(user_id, identifiers, item_type)
• Cost card administration – onboard_card(user_id, card_number, expiration_date, cvv, card_type, is_primary), request_purchase_confirmation, confirm_purchase

Conclusion

This collaboration between AWS and Visa demonstrates how agentic commerce can essentially reshape the commerce expertise, reworking what has historically been a fragmented, multistep course of right into a seamless, clever, and safe journey from discovery to buy. These capabilities signify the way forward for digital journey and buying: clever, safe, and effortlessly linked, the place trusted brokers work on behalf of shoppers to show journey intent into booked experiences in a single, unified stream. The parts in these workflows are modular and reusable throughout use circumstances within the agentic commerce ecosystem. Come be a part of the dialog and begin constructing these safe, seamless fee experiences on your prospects, utilizing Amazon Bedrock AgentCore, Strands Brokers, and Visa Clever Commerce. Right here’s the pattern GitHub repo to get began:

• Journey agent pattern: Hyperlink
• Buying agent pattern: Hyperlink

In regards to the authors

Sangeetha Bharath is a pacesetter in AI technique at Visa, the place she shapes the technical imaginative and prescient throughout developer, enterprise, and cloud segments. She focuses on neural community architectures, giant language fashions (LLMs), and reinforcement studying from human suggestions (RLHF)—experience she applies to advance AI-driven innovation in funds. Sangeetha led the event of Visa’s first MCP server and champions developer experiences that make Visa one of the best ways to pay and be paid. She additionally drives strategic development initiatives and partnerships on the intersection of AI and fintech.

Seemal Zaman is a product chief with expertise constructing and scaling FinTech merchandise. She has led zero-to-one initiatives, advanced integrations, and improvements in funds, with a present deal with making use of agentic AI to remodel B2B and client experiences. Seemal at present works on a crew centered on Visa Clever Commerce and Trusted Agent Protocol, the place she is driving innovation in agentic commerce. She thrives on the intersection of know-how and commerce, bringing daring concepts to life and turning them into merchandise that make an affect.

Isaac Privitera is a Principal Knowledge Scientist with the AWS Generative AI Innovation Middle, the place he develops bespoke agentic AI-based options to deal with prospects’ enterprise issues. His major focus lies in constructing accountable AI methods, utilizing strategies akin to RAG, multi-agent methods, and mannequin fine-tuning. When not immersed on this planet of AI, Isaac will be discovered on the golf course, having fun with a soccer recreation, or climbing trails together with his loyal canine companion, Barry.

Hardik Thakkar is a Sr. Safety Prototyping SA at Amazon Net Providers (AWS) with the Prototyping and Cloud Engineering Staff in World Monetary Providers (GFS). He focuses on safe structure design and foundations on AWS, leveraging his safety experience to serve monetary companies prospects. His focus areas embody security-first design patterns, monetary companies compliance frameworks, and serving to establishments construct strong cloud infrastructures and AI-based options on AWS.

Daniela Vargas is a Prototyping Options Architect with the Prototyping and Cloud Engineering Staff in AWS World Monetary Providers (GFS), the place she works backwards from buyer must create modern prototypes. Her experience spans from information analytics pipelines that unlock enterprise insights to cutting-edge generative AI implementations that rework buyer experiences.

Ritambhara Chaterjee is a Senior Options Architect in AWS World Monetary Providers with experience in machine studying and fee applied sciences. She helps monetary establishments innovate on the AWS Cloud by offering options for fraud detection, transaction processing, and AI-powered monetary purposes utilizing AWS services and products.

Ankit Pathak leads ML, generative AI, and agentic AI GTM apply for AWS World Monetary Providers, bringing 15+ years of technical depth throughout information, analytics, and AI engineering. His focus areas embody growing frontier agentic AI patterns together with multi-agent methods leveraging autonomous planning, tool-use optimization, secure guardrails, and long-context reasoning, mixing utilized analysis with real-world patterns to drive compliant enterprise generative AI adoption.