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 published a deployable reference architecture in which a LangChain-based enterprise orchestrator calls a Strands Agents-based compliance agent in a separate AWS account using the A2A protocol, deployed from three Jupyter notebooks. Each side has its own Amazon Cognito user pool and only public values are exchanged during onboarding, so no secrets are shared.

AWS published a reference architecture that stops a shared Bedrock agent from searching documents the signed-in user cannot see, by injecting a verified email into every tool call at the AgentCore Gateway and letting the knowledge base filter on it.

AWS's Agent DLC proposes judging every LLM agent change against an approved "gold standard" of criteria, thresholds and examples, backed by repeated-run pass rates rather than single test assertions. A reference HR policy assistant showed chunk-level retrieval precision of 1.00 while fact-level precision fell to 0.09 on one question, hiding a missing-document problem.

AWS published a method for catching silent data corruption (SDC) in PyTorch DDP training on SageMaker HyperPod: three in-loop checks that identify a suspect GPU by comparing data-parallel replicas, then cordon and replace its node. The checks were validated with software-injected faults on a cluster of 4 ml.p5.48xlarge instances, 32 H100 GPUs.

AWS will hold re:Invent 2026 from November 30 to December 4, with more than 140 sessions in the AI/ML track among over 2,000 total learning sessions, and breakouts recorded for later viewing on the AWS Events YouTube channel.

DXC Technology's OASIS agent platform, generally available since March 2026, cut incident resolution time by an average of 95% across 56 measured use cases; it runs all model calls through Amazon Bedrock behind a single API, with guardrails, private networking and approval gates layered on top.

AWS has published a catalog of 23 production-ready, open source generative AI assets from its Generative AI Innovation Center, now part of AWS Frontier AI Engineering & Services. The post cites customer results such as DoorDash handling over 100,000 calls per day and Ricoh achieving a 7-fold document processing throughput increase.

AWS published a framework called AI-PDLC that implements the first four product development phases with six Quick agents, gated by human approval at each step. The suite, its prompt templates and setup guide are available from a GitHub repository.

AWS has open-sourced UAEF, a framework-agnostic evaluation system for AI agents licensed under Apache 2.0. It normalizes traces from LangGraph, LangChain, Strands Agents, Amazon Bedrock AgentCore, Langfuse and generic JSON into a canonical AgentTrace, then scores them with 36 built-in metrics across 7 quality dimensions.

AWS launched usage-based metering for SageMaker AI bidirectional streaming, letting marketplace model providers bill by actual streamed usage via a companion metadata WebSocket, with Deepgram as launch partner for STT and TTS billing.

Moonshot AI's Kimi K3.1 Pro Highspeed is now on Amazon Bedrock, a post-trained version of Kimi K3 with a 3T-class open-weight architecture and 1-million-token context, reportedly running 2–3 times faster than Kimi K3.1 Pro. It keeps the same endpoint and APIs as Kimi K3, so existing users can switch by changing the model ID.

AWS has released the Web Search API for Amazon Bedrock AgentCore, which serves current web results from an Amazon-operated index, signed with existing AWS credentials so no search API key needs storing. LangChain's AWS integration now exposes it as a tool with fixed date, blocklist and search-budget rules.

Amazon Bedrock AgentCore Evaluations now runs AWS built-in evaluators plus managed DeepEval and AutoEval metrics against agent traces and sessions, using the same evaluator IDs across on-demand, online and batch modes. Teams can also derive custom judges from any of those bases and score several criteria in one call with multi-dimension evaluators.

Trovata has moved its treasury platform from generative AI insights to autonomous agents built on Amazon Bedrock and Bedrock AgentCore, with role-based access control and a library of ten agent patterns drawn from customer deployments. Trovata reports $1M+ in annualized savings, 5+ hours per day saved per accounting team, and reporting cut from hours to under a minute.

Tata Consultancy Services runs its Agentic Test Orchestrator in production for banking, financial services and insurance clients, turning requirements into executable tests with specialized agents hosted on Amazon Bedrock AgentCore. TCS reports an internal benchmark of roughly 100 scenarios dropping from about 200 engineer-hours to about 24, an 88 percent reduction.

AWS detailed how agents, skills and data sources fit together in AWS Quick, including that skills follow Anthropic's open Agent Skills specification and that Quick on desktop went from preview in April 2026 to general availability in September 2026.

AWS announced that Anthropic's Claude Fable 5.2 and Claude Mythos 5.2 are available on Amazon Bedrock and Claude Platform on AWS, with Fable 5.2 positioned as Anthropic's most capable generally available model for coding and scientific work and Mythos 5.2 restricted to Anthropic's verification programs for cyber security and life sciences.

NVIDIA's Nemotron 3.5 Super VL, a 120B-parameter vision-language model with 12B active parameters per token, is now deployable in Amazon SageMaker JumpStart in two variants: a default NVFP4 build on G7 instances and full-precision BF16 weights on P5en.

AWS has published the Pluggable Agentic AI Framework, a seven-layer reference architecture for running multi-agent applications in production, implemented on Amazon Bedrock AgentCore and available as a public sample repository and Workshop Studio experience.

CrossnoKaye's AURA agent runs its numerical analysis as Python inside Amazon Bedrock AgentCore code-interpreter microVMs, so plant diagnoses rest on executed statistics rather than LLM arithmetic, and results come back as cited Briefs. CrossnoKaye reports its ATLAS platform already spans more than 100 facilities and 13,000 controlled assets.