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

AWS AI Updates: November 14, 2026

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

1. Bedrock AgentCore Identity Accepts OAuth 2.0 Discovery URLs, Removing the OpenID Connect Requirement

AWS added support for OAuth 2.0 authorization server discovery URLs to Amazon Bedrock AgentCore Identity on November 13, for both inbound JWT authorization and outbound OAuth credential providers. AgentCore Identity previously accepted only OpenID Connect discovery URLs, so authorization servers that implement OAuth 2.0 without OIDC, and publish their metadata at a different well-known path, could not be used. Either kind of discovery URL now works when configuring a JWT authorizer or a custom OAuth credential provider. The feature is available in all Regions where AgentCore is offered at no additional cost. Source

2. AgentCore Runtime Adds Build Hooks, and AWS Publishes Startup Patterns for the Snapshot-Based V2 Runtime

AWS published guidance on November 13 for the V2 AgentCore runtime, which snapshots an initialized agent once per create or update and restores it for each session. The post states that the runtime now supports build hooks: an HTTP endpoint in the container that AgentCore calls once before the snapshot and once after a test restore, with the two calls sharing 120 seconds and a failure failing the deployment. It reverses the usual cold start advice, recommending that imports, clients and static data be created during initialization because work deferred to the handler is paid by every new session. It also warns that random state, timestamps and self-fetched credentials created during initialization are identical in every session, which it calls a security issue for random values. For chat and voice, the post suggests starting the session in the background when the user opens the app, since a new session takes about 2 seconds and a running one answers in about 200 ms. Source

3. SageMaker HyperPod Data Capture Records Live Inference Traffic to S3 at Three Tiers

AWS described Data Capture for Amazon SageMaker HyperPod on November 13, a built-in capability that records inference inputs and outputs to Amazon S3. It has three independent tiers: full payloads at the SageMaker endpoint for compliance and Model Monitor, Application Load Balancer access logs for traffic metadata, and full payloads plus event metadata at the model pod, written by sidecars that the HyperPod inference operator adds. The walkthrough uses the pod tier on a support ticket router, finds with Amazon Athena that tickets for a newly launched product are being routed to a catch-all class, labels the captured records with a SQL policy and fine-tunes with the HyperPod training operator. On 280 held-out tickets, accuracy rose from 235 to 280 correct with no regression on the five existing classes; the data is synthetic. The post advises sampling, AWS KMS encryption and PII redaction because the pod tier stores complete payloads. Source

4. OpenFold3-preview Becomes Deployable From SageMaker JumpStart on Asynchronous Endpoints

AWS announced on November 13 that OpenFold3-preview, the Apache 2.0 biomolecular structure prediction model from the OpenFold Consortium and the AlQuraishi Lab at Columbia University, can be deployed from Amazon SageMaker JumpStart. It predicts the joint 3D structure of proteins, DNA, RNA and small-molecule ligands, returning mmCIF files with pLDDT, PAE and PDE confidence metrics. Because folding often takes minutes, the model runs on SageMaker Asynchronous Inference, which queues requests from S3, allows invocations of up to 1 hour and can scale to zero when idle; the endpoint runs with network isolation. Users can supply their own multiple sequence alignments, and in the post’s example a protein and ligand complex went from an average pLDDT of 30 without an alignment to 98 with one. The developers describe the release as a research preview. Source

5. Grok 4.7 Arrives in Kiro for AWS GovCloud (US) With a 500K Context Window

AWS made Grok 4.7 available in the Kiro IDE and CLI for the AWS GovCloud (US) Regions on November 13, with experimental support and a 500K context window. Requests up to 200K tokens use a 2.7x credit multiplier, and requests above 200K tokens are billed at 5.4x. The announcement attributes the model to “SpaceXSI” and describes it as built for long-running coding and knowledge work, with more reliable planning and error recovery than Grok 4.6 on repository-scale tasks. Kiro administrators control which models their users can select, and the IDE or CLI must be updated and restarted for the model to appear. Source

6. SageMaker Unified Studio Lets Administrators Set Which Tools Each Designation Sees

AWS added layout customization to Amazon SageMaker Unified Studio on November 13. Domain administrators define a layout inside a designation, and it applies in every project in the domain, so a SQL analyst designation can show only the Query Editor while an ML engineer designation shows notebooks, training jobs and model endpoints. Configurable navigation items span Data Analytics, AI/ML, Generative AI and IDEs. A member who opens a hidden page directly gets an access-denied page, members with several designations see the union of their layouts, and navigation items AWS adds later stay hidden for designations with a configured layout until an administrator adds them. The feature is available in all Regions where SageMaker Unified Studio is offered. Source

7. AWS Sample Routes Coding Agent Requests Across Bedrock Models With NVIDIA NeMo Switchyard and LiteLLM

AWS published a CDK sample on November 13 that deploys LiteLLM with the open-source NVIDIA NeMo Switchyard routing plugin on Amazon ECS on AWS Fargate, calling models on Amazon Bedrock. Stage routing reads the tool activity and results in the request history to choose between a capable model and a cheaper one, and Random routing splits traffic by weight for A/B tests; neither calls a judge model. The tested pair is Claude Opus 5.5 and NVIDIA Nemotron 3 Super 120B in US East (N. Virginia). The sample’s benchmark runners report cost per solved task, defined as the total cost of all runs divided by the runs whose tests pass, for the routed group against single-model baselines. The post publishes no benchmark results, and labels the Switchyard LiteLLM integration experimental and the sample as non-production code. Source