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
AWS launched its first brand campaign for AWS Quick, an AI workspace built only for enterprise work, running in the US and UK across streaming TV, online video, social media and business publications. The product is now branded AWS Quick to signal enterprise security and governance.

Amazon SageMaker Inference now supports NVIDIA Multi-Instance GPU partitioning, letting you split a physical accelerator into up to seven hardware-isolated partitions on inference component endpoints. Each partition gets dedicated compute and memory, so several models can share one GPU without interference.

AWS's Quick team published a taxonomy that breaks agentic work into four primitives — triggers (When/What), agents (Who), and capabilities such as tools, skills and workflows (How). In their framing, a request like a daily Slack digest is usually a trigger attached to an existing agent, not a new agent.

Amazon's North America Sort Center team built a serverless Model Context Protocol (MCP) server on Amazon ECS Fargate that gives LLM agents a shared gateway to its data lake, where adding a dataset requires only a JSON metadata file rather than custom code and a redeploy.

AWS published a walkthrough for letting an AI agent act on behalf of signed-in users by combining Okta Cross App Access with Amazon Bedrock AgentCore runtime and AgentCore Identity. In AWS's tests, removing a user's assignment to the agent blocked new requests in about 1 second, while already-issued tokens expired within 5 minutes.

Greenlight Financial's Luna system, built on Amazon Bedrock AgentCore, autonomously resolves more than 90 percent of its 3,000-4,000 monthly security and risk cases, cutting routine triage from 15-30 minutes to seconds. It uses a two-tier supervisor plus over 30 specialized agents and episodic memory.

AWS has made Web Answer generally available on Amazon Bedrock AgentCore, a managed AgentCore Gateway tool that takes a natural-language question and returns a cited answer in a single call, priced at $10 per 1,000 tool calls in us-east-1.

Amazon described a content-quality evaluation system for A+ Content that fuses Vision-Language Model scores with OCR readability signals in a CatBoost classifier trained on expert labels, reaching 0.94 accuracy and 0.95 F1 on a holdout set, up from a 0.80/0.84 baseline.

AWS Quick now has Smart Approvals, which uses a separate classifier model to automatically approve routine agent actions while sending consequential ones, such as external sends, access changes, deletions and payments, back to the user. It is on by default, cannot exceed existing permission controls, and each approval is single-use.

AWS published a sample implementation that uses the open source Python crawler Crawl4AI to prepare web pages as Markdown plus metadata sidecars in Amazon S3, which Amazon Bedrock Knowledge Bases then ingests through its S3 data source connector. It targets sites needing tighter control than the managed Web Crawler offers.

AWS published a walkthrough for distilling a multi-step shopping agent from a frontier teacher model (GLM-5.3 on Amazon Bedrock) into a Qwen3.5-4B student using sequence-level knowledge distillation and serverless supervised fine-tuning on SageMaker AI. The recipe generates two teacher trajectories per training problem and keeps only runs the ShoppingBench scorer marks as solved.

AWS published a walkthrough of a multi-agent engineering assistant on Amazon Bedrock AgentCore that gathers part-substitution evidence using Amazon Nova Multimodal Embeddings, deterministic Python scoring, and the open source Strands Agents SDK, while leaving the approval decision to the engineer.

Circle K cut mean incident triage time from 45–60 minutes to under 15 minutes using an agentic system built on Amazon Bedrock, now running in production on Bedrock AgentCore Runtime after an earlier version on Bedrock Flows.

AWS published a five-stage operating model for controlling Amazon Bedrock spend, built on invoice-grade per-identity cost attribution through IAM principals and CUR 2.0. The playbook targets FinOps and platform teams, with three enforcement tiers ranging from budget alerts to gateways that block over-budget requests synchronously.

AWS published a deployment guide showing how to point the ChatGPT desktop app and its Codex coding agent at Amazon Bedrock in AWS GovCloud us-gov-west-1, so prompts and responses stay inside the selected region under IAM and network controls. Billing is per token with no seat licenses, and zero data retention is supported.

The UK Amazon Quick User Group held three quarterly meetups in 2026, all at Amazon's Shoreditch office in London, drawing more than 150 attendees in total to hands-on workshops covering custom chat agents, Quick Desktop, Quick Apps, Flows, Automate and Research.

Epsilon's 50-engineer Agentic Platform team built a multi-tenant UI testing service on Amazon Nova Act and Bedrock AgentCore Browser Tool, reporting 4–6x faster test authoring (5–10 minutes versus 30–90 for an equivalent 18-step Playwright script) and a 100 percent pass rate on replayed trajectories.

MoEngage wired roughly 400 private REST APIs across seven AWS Regions into agent tools using Amazon Bedrock AgentCore Gateway, avoiding more than 20 separate MCP servers and estimating about six months of engineering saved. New APIs now onboard in hours via a spec annotation rather than days of networking and review.

NVIDIA's Nemotron 3.5 Super VL, a 120B-parameter open multimodal mixture-of-experts model with 12B active parameters per token and up to 1M-token context, can now be deployed on Amazon EKS, EKS Auto Mode and SageMaker HyperPod EKS using a reference implementation with NVIDIA Dynamo and vLLM.

AWS's Ray Train Deep Learning Container, published on Amazon ECR Public, packages Ray 2.58.0, PyTorch 2.13.0, CUDA 13.0.2, NCCL 2.29.7 and the EFA networking stack so teams can run multi-node PyTorch FSDP training on EKS without building or patching an image.