AI news for builders and product teamsUpdated Nov 17, 2026, 13:01 UTC
AWS Machine Learning Blog
First-party releases and research from AWS Machine Learning Blog. Headlines and excerpts link to the original articles.
Latest stories
Newest first
Octave built a shared multi-tenant AI foundation on Amazon Bedrock for its Reliance quality management platform, using Cohere Embed v4 and Bedrock AgentCore to embed 2 billion documents and power a Complaint Triage Assistant now running in production. The company says the foundation cut deployment time per use case by 50 percent.

Amazon Bedrock now serves OpenAI's GPT-5.6, GPT-6 and GPT-6.1 Sol models to applications running in the Canada (Central) and Canada West (Calgary) regions via cross-Region inference, which routes requests to US or global capacity pools depending on the chosen profile.

AWS Agent Registry, now generally available, is a managed catalog of MCP servers, agents and agent skills that an orchestrator can search at runtime so new capabilities are usable after registration and approval rather than a code change and redeploy.

AWS and Keyfactor published a joint reference architecture that gives AI agents verifiable X.509 identities, using signed request envelopes, user-signed delegation, RFC 8693 audience-scoped tokens and real-time OCSP revocation checks. It targets the authentication gap in MCP, which defines no built-in agent identity standard.

Eon replaced per-tool credentials and local MCP servers with a single Amazon Bedrock AgentCore MCP Gateway, cutting new-engineer onboarding from roughly two days to about two minutes. The gateway serves 100+ developers and an autonomous SRE agent across AWS, GCP and Azure with Cedar policy enforcement.

A new AWS and Denodo walkthrough shows a Strands agent on Amazon Bedrock AgentCore querying a Denodo semantic layer over Aurora, Redshift and an on-premises PostgreSQL cluster, with joins, metric definitions and row/column filters enforced in the layer rather than the prompt.

AWS now lets new SageMaker AI customers sign up with Google, GitHub, Apple, Amazon or email and land in a preconfigured SageMaker Studio in 2-3 steps, with $100 in Free Tier credits and no credit card for most.

Amazon SageMaker JumpStart onboarded more than 100 new models from 30 providers between January and September 2026, with Alibaba's Qwen line contributing 29. AWS also added a Model Hub on the SageMaker AI console, four named deployment profiles, and a Hugging Face deep link that opens a pre-configured Studio environment.

TwelveLabs' Pegasus 1.5 video language model is now on Amazon Bedrock, adding time-based segmentation that returns every matching video segment as timestamped JSON, plus multimodal prompting with reference images and support for videos up to two hours in a single request.

Deepgram and AWS ported the Nova-3 speech-to-text model to Trainium2 for offline batch workloads, running eight model replicas on one trn2.3xlarge instance and processing 11–70 percent more audio per dollar than the compute options Deepgram had been running.

AWS contributed a push-based NIXL KV connector to vLLM that overlaps decode-side setup with prefill, reporting up to 3x better time-to-first-token and about 30 percent better time-per-output-token versus the default pull mode on a Qwen3-32B benchmark.

AWS's Sales, Marketing, and Global Services Operations team described a nine-stage evaluation framework built on Amazon Bedrock AgentCore Evaluations for its production multi-agent sales intelligence system, after pre-launch testing caught silent errors such as a three-month figure of $410.08 reported as "$410.08 Million."

Reflection AI released Beam, an open-weight sparse mixture-of-experts model for agentic coding with 501 billion total parameters and 23 billion active parameters, now deployable on Amazon SageMaker JumpStart and HyperPod Inference.

Meta's Muse Spark Open, an 816B-parameter mixture-of-experts coding model with a 1M-token context window, is now deployable through Amazon SageMaker JumpStart in BF16 and NVFP4 quantized variants. It runs as an endpoint in the customer's own AWS account and is available at launch in US East (N. Virginia) and US East (Ohio).

AWS introduced Trust in Depth, a framework that layers agent scoping, identity, runtime policy enforcement and trace-based observation so enterprises can grant agents autonomy where rules can be machine-verified and keep humans on judgment calls.

AWS published a reference architecture for a self-improving intelligent document processing system built on Amazon Bedrock AgentCore, in which eight specialized agents coordinate across training and automation workflows to classify, extract, and validate business documents. The sample deploys via AWS CDK in roughly 15–20 minutes, runs about $58 per month for 10 training and 1,000 processed documents, and processes repeat document types faster after learning extraction rules.

AWS published a walkthrough and open-source sample for an agent system on Amazon Bedrock AgentCore that turns enterprise documents into PowerPoint decks, which completed in 70-90 seconds in internal testing with model costs around $0.04 per generation.

Tamnoon built Tami, an agentic cloud remediation orchestrator on Amazon Bedrock, which autonomously investigates security findings, scores whether fixes are safe, and generates remediation code with guardrails. AWS reports the approach drove over 90 percent reduction in critical cloud exposure in 90 days with zero production issues.

Anchor Browser's Agentic Payments, built with the AWS Generative AI Innovation Center Partner Agent Factory and now on AWS Marketplace, lets AI agents pay for x402-protected pages and APIs automatically through Amazon Bedrock AgentCore payments. Enabling a single session property, session.x402, is all that's required.

AWS published a sample repository showing two ways to activate AWS Marketplace Private Pricing Offer entitlements for Amazon Bedrock in AWS License Manager: a one-time script and an event-driven CDK stack. Organization-level grants are auto-accepted in AWS Organizations but stay Disabled until activated, so negotiated pricing does not apply until someone acts.