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's Agentic Retrieve on Amazon Bedrock Managed Knowledge Base now accepts a memoryConfiguration field that makes AgentCore memory a retrieval source alongside the knowledge base, so conversation history and extracted user facts steer searches within the same call. The integration is opt-in and backward compatible, and the same request passes a userContext so document ACLs still filter results.

Duco rebuilt its agentic AI stack on Amazon Bedrock and Bedrock AgentCore, using per-tenant runtimes and an MCP skills architecture that shrank context from roughly 200,000 tokens to about 20,000. It reports reconciliation deployments 10x faster and, for one global bank, 70% lower onboarding costs.

AWS's ML blog lays out a two-tier taxonomy of multi-agent design patterns, arguing that coordination failures — not models — sink most pilots, and that teams should start with deterministic workflows and only move to model-chosen control flow when subtasks emerge during the work.

MoEngage built its Merlin AI agent layer on Amazon Bedrock AgentCore by turning existing OpenAPI 3.0 API specifications into MCP tools, gaining tenant-isolated sessions and reusing JWT authentication without rebuilding its APIs. The company reports customers like Glance cut campaign go-live time 50% and lifted push click-through rates 38%.

AWS published a five-phase methodology for migrating mainframe Assembler batch programs to Java Spring Boot on AWS, using Amazon Bedrock for business-rule extraction and code generation, an intermediate COBOL stage for SME review, and AWS Transform for the final COBOL-to-Java conversion.

Novo Nordisk's Manufacturing Technology team in the Tablet area built a GraphRAG-based AI assistant on AWS, pairing a Neptune knowledge graph with OpenSearch Serverless hybrid search and agents on Amazon Bedrock AgentCore, reducing some SME work from hours or days to minutes.

Amazon Bedrock Managed Knowledge Base now includes tabular search, which runs model-written code over whole CSV, XLSX and XLS files in an isolated sandbox so agentic retrieval can answer questions requiring calculations across every row, not just the passages retrieval returns.

AWS's guide cautions that Bedrock Guardrails Denied Topics match by meaning, so domain-named topics like Finance or Legal block legitimate requests; it recommends naming topics after user intentions, writing action-scoped definitions, and using varied boundary examples.

Amazon Bedrock Managed Knowledge Base now supports retrieval instructions, a managed, versioned resource that lets one shared knowledge base answer in a per-tenant voice, language, format and policy, pinned by version per request. Instructions shape answers but are not an authorization boundary; document access still needs metadata filters or separate knowledge bases.

Open-weight models on Amazon Bedrock now support explicit prompt caching, letting developers mark reusable prompt prefixes with cache breakpoints so repeated context is read from cache instead of reprocessed. Implicit caching remains available and needs no request changes.

Amazon Bedrock AgentCore runtime's V2 platform version snapshots a prepared microVM once at create or update time and restores it for each new session, cutting and stabilizing session start times. AWS also added build hooks so containers can prime large files, CLIs and browsers during snapshot creation.

AssemblyAI runs its speech-to-text inference on Amazon EC2 Spot GPU instances across g4dn, g5 and g6 families, reporting 70 percent cost savings on inference compute versus On-Demand pricing while holding SLAs. A queue-based retry design absorbs Spot interruptions, and independent regional deployments in us-west-2 and eu-west-2 support data residency.

Amazon SageMaker HyperPod now has Data Capture, which records live inference requests and responses to S3 at three independent points in the request path, with minimal latency impact and no effect on model availability.

Amazon announced that SGLang-Neuron, released by RadixArk with Annapurna Labs, is publicly available through an AWS Deep Learning Container, letting teams serve models and run reinforcement learning rollouts on AWS Trainium through the standard SGLang interface.

Natera built SWARM, an agentic AI system on Amazon Bedrock AgentCore, that cuts average first response time on lab equipment support tickets from 17.4 hours to roughly 8 minutes (about 130x faster) and gives 96.2 percent of tickets an automated first response.

An AWS post details how four request-level Amazon Bedrock settings — prompt cache TTL, Flex service tier, adaptive thinking effort and maxTokens — change token counts and rates without touching model or prompt. In its Opus 5 agent test, low effort used 936 median output tokens versus 2,521 at the default high.

AWS published a no-code pattern for attributing Amazon Bedrock costs to individual IAM principals using automatic identity tagging, cost allocation tags, CUR 2.0 or Cost Explorer, and AWS Budgets actions that can attach a Deny policy when a threshold is crossed.

A new AWS Bedrock framework argues that comparing models by token price is misleading: in a worked example, a model priced twice as high delivered the same agentic task for $0.00145 versus $0.0042, about 65% cheaper.

AWS's ML blog describes a self-service LLM gateway pattern that puts LiteLLM on Fargate in front of Amazon Bedrock, exposing an OpenAI-compatible API with per-user keys, spend caps and observability. Note that self-service key issuance needs the licensed LiteLLM Enterprise edition; the open source gateway is documented as free for up to five users.

AWS added a knowledge base migration guide to the amazon-bedrock skill in the Agent Toolkit for AWS, letting a coding agent inspect an existing Amazon Bedrock customer-managed knowledge base read-only, map its configuration, and optionally build a MANAGED replacement alongside it. Ingestion, application changes and cutover stay with the team.