AWS AI Updates: November 6, 2026

1. Amazon Bedrock AgentCore Launches Web Answer, a Managed Tool That Returns a Cited Answer in One MCP Call for $10 per 1,000 Calls

Amazon Bedrock AgentCore made Web Answer generally available, a managed tool that takes a natural-language question and returns a synthesized answer with citations to current web sources, so agent builders do not have to run their own search, retrieval, synthesis and citation pipeline. It extends AgentCore Web Search, which returns ranked URLs, titles and snippets for the customer’s own model to reason over; Web Answer instead runs the model and an agentic loop that plans searches, runs several in parallel, reads key pages to check facts and searches again when evidence is insufficient. It attaches as a connector target (connectorId: "web-answer") on an AgentCore Gateway that uses IAM or JWT inbound authentication, and agents discover and call it over MCP through tools/list and tools/call; the gateway’s service role needs bedrock-agentcore:InvokeWebAnswer on arn:aws:bedrock-agentcore:us-east-1:aws:tool/web-answer.v1, and outbound credentials are handled by the GATEWAY_IAM_ROLE provider. Administrators can set include and exclude domain lists on the target, which agents can narrow per call but not widen, and each call can also pass a published-date range and a country to localize results. Responses carry markdown text with inline citation markers plus a structuredContent.answer block listing each citation’s URL and domain (the sample response reports 5,210 ms latency). AWS says Web Answer searches an Amazon-operated web index and Amazon’s knowledge graph without sending questions to third-party search engines, supports PrivateLink through the gateway VPC endpoint, publishes CloudWatch metrics and enforces a default limit of 10 transactions per second per account. It is available only in US East (N. Virginia), priced at $10 per 1,000 tool calls. Source Source

2. Amazon OpenSearch Service Optimized Engine Stores Agent Traces, Prompts and Tool Calls Without Flattening

Amazon OpenSearch Service added agent trace support, including OpenTelemetry traces, to its optimized engine, so traces and logs for a request or an agent session sit in one place and the failing or slow step can be found next to its related logs. The engine maps OpenTelemetry logs and traces to object, nested and flat object fields with support for multi-valued arrays, and span events, span links, prompts and tool calls keep their nested structure, which allows field-level queries without flattening data at ingestion. On hot-tier data it supports update, delete and upsert by document ID, so late or erroneous records can be corrected in place for dashboards and alerts. OpenSearch UI provides Discover Traces, application performance monitoring, a service map and dedicated agent traces and sessions views. The features require OpenSearch 3.7, carry no additional charge and are available in 12 Regions: US East (N. Virginia, Ohio), US West (Oregon), Canada (Central), Asia Pacific (Mumbai, Singapore, Sydney, Tokyo) and Europe (Frankfurt, Ireland, London, Spain). Data can be sent with Amazon OpenSearch Ingestion. Source

3. Amazon Q Developer in the AWS Console Is Renamed the AWS Assistant

AWS renamed Amazon Q Developer in the AWS Management Console and the Console Mobile Apps to the AWS assistant, keeping the same personalized, account-aware guidance and actions under a name AWS says communicates what the product does. The AWS assistant is currently free, existing Amazon Q Developer Pro subscribers keep their Pro tier capabilities automatically, and it is available globally, with a new product page at aws.amazon.com/products/assistant. The announcement covers only the console and mobile app surfaces and does not say whether the Amazon Q Developer IDE extensions or CLI change. Source

4. AWS WAF Lets Publishers Take AI Agent Payments as Local-Currency Stripe Payouts

AWS WAF AI traffic monetization now integrates with Stripe, so publishers, content owners and API providers can collect stablecoin payments from AI bots and agents and have them settle in local currency to their Stripe balance, without holding stablecoins or managing a crypto wallet. Stripe is added as a payout method alongside the existing stablecoin payouts and is connected through Stripe Connect in the AWS WAF console; existing pricing and agent policies keep working unchanged. When an agent requests protected content, AWS WAF returns an HTTP 402 response listing the payment credential types it accepts, and the content is served once the agent retries with a valid credential. Revenue from all payment providers appears in the console’s revenue analytics. The integration has no additional charge for AWS WAF customers using Amazon CloudFront in all edge locations where web ACLs are associated with CloudFront distributions, while standard AWS WAF and Stripe processing fees apply. Source

5. Amazon Fuses VLM Rubric Scores With OCR Readability in a CatBoost Classifier to Grade Product Pages, Reaching 0.94 Accuracy

Amazon applied scientists and AWS architects described how Amazon rates A+ Content (the rich media modules on product detail pages) by rendering each page as a single snapshot and scoring it with a frozen, unnamed vision-language model against a rubric of four dimensions (Visual Storytelling, Feature Presentation, A+ Layout, Conversion Optimization), each returning a 0 to 10 score, a confidence and a rationale. Instead of fine-tuning the VLM, a CatBoost classifier is trained on those scores plus four OCR readability signals (text density, character density, spacing, overall readability) against labels from one ecommerce expert and four annotators on 5,478 pages across 10 categories (49.6% high quality; each annotator scored 0.845 to 0.915 Cohen’s kappa against the majority vote). On a 20% holdout of about 1,100 pages, averaging the VLM scores with a fixed threshold reached 0.80 accuracy and 0.84 F1, OCR-only CatBoost 0.82 and 0.83, VLM-only CatBoost 0.89 and 0.90, and the fused model 0.94 and 0.95; on a 499-project single-category set, fusion lifted a lighter VLM from 0.74 to 0.90 accuracy and a more capable one from 0.79 to 0.94. SHAP values rank OCR character density (1.1), overall readability (1.0) and Visual Storytelling (1.0) as the most influential features and feed an LLM that writes prioritized fixes; pages revised with those recommendations gained 1.4 points on the 10-point quality scale, and a worldwide A/B test that changed the visibility of low-quality content produced what the authors call a statistically meaningful lift in Ordered Product Sales, without a figure. The SageMaker AI implementation uses Processing Jobs for rendering and SHAP, real-time, asynchronous or batch inference for the VLM, a JumpStart OCR model, Feature Store, Training Jobs, Model Registry and Pipelines for drift-triggered retraining, and a diffusion model that generates improved image assets. Source

Amazon SageMaker AI now supports IAM Identity Center multi-Region replication, so a SageMaker AI Studio domain can be deployed in a Region other than the primary Identity Center instance’s Region by replicating that instance into the Region where the domain runs; previously both had to be in the same Region. AWS positions the change for organizations that keep centralized workforce identity and single sign-on in one Region while running Studio workloads in Regions required for data residency or compliance. It is available in all standard Regions where SageMaker AI Studio is supported, including US East (N. Virginia, Ohio), US West (Oregon), Canada (Central), South America (São Paulo), Europe (Ireland, Frankfurt, London, Paris, Stockholm) and Asia Pacific (Tokyo, Seoul, Singapore, Sydney, Mumbai). Source