Monday, November 16, 2026 Edition No. 240
Search
Daily News Fully automated · 7 min read

AWS AI Updates: November 13, 2026

Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled

1. Next-Generation AgentCore Runtime Becomes the Default, With Sub-2-Second Session Starts and Scale to Zero

AWS made the next generation of the Amazon Bedrock AgentCore runtime the default for new runtimes in 19 Regions on November 12. Each session is restored from a snapshot of the initialized agent inside its own hardware-isolated Firecracker microVM, scales up to 8 vCPU and 32 GB of memory, and supports container images up to 10 GB. AWS reports session starts under 2 seconds at P75 on a 2 GB image, down from 29.7 seconds, and up to 68% less billed memory per session for multi-turn agents because idle agents scale to zero. Existing runtimes stay on the current version until platformVersion is set to V2, and the accompanying blog post adds committed baseline pricing for always-on agents and the ability to mount a filesystem such as an Amazon S3 bucket inside the session. Source

2. SGLang-Neuron Brings the SGLang Serving Stack to AWS Trainium Through a Deep Learning Container

AWS announced on November 12 that SGLang-Neuron, built by RadixArk with Annapurna Labs, is publicly available as an AWS Deep Learning Container. It is an out-of-tree SGLang plugin selected with SGLANG_PLATFORM=neuron: SGLang core keeps the API server, RadixAttention and prefix-cache bookkeeping, while the plugin supplies Trainium-specific models, attention, Mixture-of-Experts support and a graph runner built on TorchNeuron, Neuron Kernel Interface kernels and torch.compile. Prefill and decode disaggregation is supported, with the KV cache transferred between servers using NIXL over Elastic Fabric Adapter. The walkthrough serves openai/gpt-oss-120b on a Trainium3 EC2 instance behind the standard OpenAI-compatible chat completions endpoint, and the post publishes no throughput or cost figures. Source

3. Amazon Cognito Adds OAuth 2.0 Token Exchange So Agents and MCP Servers Can Call Downstream Services With Scoped Tokens

AWS added OAuth 2.0 token exchange (RFC 8693) to Amazon Cognito on November 12. A service or AI agent acting for a user can trade the user’s token for a new one that has a narrower scope, is bound to a single audience and carries no refresh token. When an actor token is included, Cognito adds an act claim identifying the intermediary, so a backend API can distinguish a call made by an MCP server on a user’s behalf from one the user made directly. The exchange is available through the token endpoint or the new GetDerivedToken API on the Essentials and Plus feature plans in all Regions where Cognito is offered. Source

4. EC2 Image Builder Console Gains AWS DevOps Agent Chat for Diagnosing Up to 50 Failed Pipelines at Once

AWS integrated the AWS DevOps Agent into the EC2 Image Builder console on November 12. The agent reads build logs, workflow steps, components and network configuration, then returns a root cause and remediation steps, and it can diagnose up to 50 failed image pipelines in one conversation. It can also create resources from a description, populating the console form with validated values for the user to review and submit. The same capabilities are exposed to customer-built agents through a new EC2 Image Builder skill in the AWS Agent Toolkit, and the feature is available in all commercial Regions with DevOps Agent pricing. Source

5. AWS Shows How to Stop an LLM From Choosing Whose Permissions a Managed Knowledge Base Search Runs Under

AWS published a design on November 12 for per-user document access when Amazon Bedrock Managed Knowledge Base is exposed to agents as MCP tools through AgentCore Gateway. The post points out that the knowledge base’s ACL awareness filters results by the userContext in the request, but once the search is an MCP tool that field becomes a tool argument the model fills in, and neither the gateway nor the knowledge base checks it against the authenticated user. The sample therefore injects identity-provider-verified user attributes into tool calls at the gateway layer and scopes the search by those attributes at the knowledge base, with additional independent layers in case that mechanism is bypassed. The stated threat model is prompt injection or model error causing a shared agent to search as a different employee or tenant. Source

6. Agent Toolkit for AWS Adds a Guide That Lets Coding Agents Plan Migrations to Bedrock Managed Knowledge Base

AWS added a knowledge base migration guide to the amazon-bedrock skill in the Agent Toolkit for AWS, described in a November 12 post. A coding agent starts with read-only discovery of the customer-managed knowledge base and produces a plan for line-by-line review; after approval it creates the new Managed Knowledge Base and data sources alongside the original, or writes a migrated infrastructure-as-code template. Ingestion and the change to the application’s Retrieve calls remain with the team. AWS recommends the managed option because Bedrock runs ingestion, storage and indexing with always-on hybrid search, a managed embedding model and reranker, native connectors for sources such as S3, SharePoint and Confluence, and ACL-aware retrieval. Source

7. AWS Describes Capping Bedrock Spend per IAM Principal Using CUR 2.0 and AWS Budgets, Without Custom Code

AWS published a walkthrough on November 12 for per-identity Amazon Bedrock cost reporting and spend limits built only from native features. Bedrock records the IAM user or role for every inference request automatically, and teams that share one role can add per-user detail with session tags passed at role assumption. Reporting comes from Cost and Usage Reports 2.0 or Cost Explorer, and enforcement uses AWS Budgets with budget actions that restrict Bedrock access when a threshold is crossed. The post positions this as the lower-maintenance alternative to the Athena and Lambda enforcement loop described in an earlier Jamf case study, which gives nearer-real-time control at the cost of owning the code. Source

8. Natera Reports 130x Faster First Response on Lab Instrument Tickets From an AgentCore-Based System

AWS and Natera, a cell-free DNA testing company, described SWARM (Systematic Workflow Analytics and Resolutions Manager) on November 12, an agentic system for lab instrument support built on Amazon Bedrock AgentCore. Before SWARM, the service engineering team handled more than 65 tickets per day across three Jira queues with an average first response time of 17.4 hours. Natera reports about 130x faster first response and 96.2% automated coverage of support tickets, measured from Jira ticket metadata over a 90-day production evaluation period. Diagnostic accuracy is more modest: tickets the system labels High confidence reach 74% accuracy across 154 evaluated cases, against 65% for Moderate and Low confidence tickets, and adding missing vendor manuals lifted the lowest-scoring instrument family from 71% to 75% with no model change. Source