Anil Nadiminti is a cloud architect whose work connects enterprise infrastructure with the capabilities and controls that let AI agents act on a user’s behalf. A senior solutions architect at AWS at the time of AI Engineer World’s Fair 2026, he has worked across cloud modernization, location-aware tools, secure AI agents, and machine-to-machine payments.
From cloud modernization to agent capabilities
Earlier in his career, Nadiminti was a senior IT architect at Educational Testing Service. His account of public-cloud adoption at ETS addressed the operational demands of delivering assessments worldwide: placing services closer to test takers, expanding computing capacity during assessment windows, and using managed and serverless services to reduce infrastructure work. He treated modernization as an organizational change too, requiring cross-functional governance, staff education, and opportunities to build cloud-native applications.
At AWS, his customer work has included enterprise financial institutions and blockchain-native companies, spanning cloud infrastructure, Web3, and decentralized finance. His work on agent tools adds another dimension: he collaborated with Sébastien Couturier on the Amazon Location Service MCP Server, incorporating it into the AWS MCP repository. The server exposes place search, geocoding, and routing through the Model Context Protocol, giving agents tools for finding places, resolving locations, and calculating routes. His public announcement of the collaboration connects his infrastructure background with the practical task of making existing services usable by agents.
Giving agents capabilities—and governing their use
In 2024, Nadiminti co-authored an insurance-claims application design combining Amazon Bedrock Agents with Amazon Verified Permissions. The example illustrates why a conversational interface must preserve existing access rules. An administrator could list claims across regions without reading individual records, while an adjuster could read and update assigned claims within their region. Authorization depended on the user’s role, record assignment, and business region; asking the agent in natural language did not grant additional entitlements.
His co-authored guidance on Policy in Amazon Bedrock AgentCore develops this approach through deterministic policy enforcement before tool execution. In a healthcare scheduling example, the gateway checks whether the patient identifier in a tool request matches the authenticated user’s identity. It can reject a request for another patient’s record even if the model chooses that action. Cedar policies make these rules explicit and auditable: access is denied unless permitted, and explicit prohibitions take precedence. The model chooses which tool to request; authorization logic determines whether that request may proceed.
Nadiminti applies a similar separation to agentic commerce in his account of building x402 applications on AWS. He starts with a familiar failure: an agent discovers useful information behind a paywall and stalls while a person arranges a subscription, enters credentials, or makes a payment. Publishers face their own tradeoff. Blocking automated access can limit discovery and licensing opportunities; allowing unrestricted access can increase infrastructure costs without compensating them for the content.
Three mechanisms make his approach concrete:
Payment as a request credential. Nadiminti explains x402 payments as a way to purchase individual resources through HTTP. A server responds with HTTP 402 payment requirements, the client supplies payment authorization, and a facilitator verifies and settles the transaction before the server returns the resource. This lets an agent obtain a piece of information or use an API without arranging an ongoing subscription. He emphasizes that conventional transaction fees can overwhelm the value of very small purchases, making payment economics central to the design.
Spending controls outside the reasoning loop. On the buyer side, he presents AgentCore Payments as a separate layer for wallet integration, payment orchestration, and spending limits. Developers can create payment sessions with maximum spending amounts and expiration times. Imported wallet keys remain in protected storage and are inaccessible to the agent. When a resource request encounters a payment requirement, the payment layer handles authorization and settlement before the agent continues. Separating payment execution from open-ended reasoning addresses the risk that malicious content or poisoned instructions could steer the agent’s behavior, while allowing it to make purchases within defined limits.
Pricing by path, identity, and intent. On the seller side, he describes AWS web application firewall controls that apply pricing at the edge. A publisher could charge differently for a blog and a research endpoint, offer different terms to verified partner bots, or distinguish model-training requests from requests serving a search. Bot detection, verification, and traffic analysis inform those rules; revenue dashboards help publishers see which bots and content paths generate payments. Applying the controls at the edge lets publishers introduce paid access without rebuilding their content origins.
Across these contributions, Nadiminti addresses both what agents can do and the conditions under which they should do it. Location tools supply concrete capabilities; authorization policies constrain access to records; payment controls bound spending; and publisher rules establish terms for automated access. His work brings the operational concerns of enterprise cloud systems into applications where models increasingly choose the next action.
Anil Nadiminti explains how x402 turns a payment into access to a resource, how AgentCore Payments separates spending from the agent loop, and how AWS WAF lets sellers price AI requests at the edge.
x402 turns a resource request into a payment exchange: the server returns Payment Required, receives authorization, uses a facilitator for verification and settlement, then delivers content.
AgentCore Payments places session budgets, expiry, wallet integration and payment execution outside the agent's reasoning loop. Imported private keys remain in a KMS-protected store the agent cannot access.
The economic target is payment overhead small enough for individual resource purchases. A 25-cent fixed fee is 250 times a tenth-of-a-cent purchase; the closing examples extend these purchases to inference, compute and MCP tools.
▶ Watch ↗