---
title: "Oct 24–30: Bedrock adds web search controls, and more"
show: "AWS Drama"
canonical: "https://podcasts.robotpaper.ai/aws/ep/2026-10-30-week/"
date: "2026-10-30T03:44:40+00:00"
duration: "46:38"
duration_seconds: 2798
audio: "https://podcasts.robotpaper.ai/aws/ep/2026-10-30-week/2026-10-30-week-aws.mp3"
chapters: "https://podcasts.robotpaper.ai/aws/ep/2026-10-30-week/chapters.json"
transcript: "https://podcasts.robotpaper.ai/aws/ep/2026-10-30-week.txt"
feed: "https://podcasts.robotpaper.ai/aws/feed.xml"
language: "en"
author: "Roy Osherove"
ai_generated_voices: true
---

# Oct 24–30: Bedrock adds web search controls, and more

AWS Drama · Oct 30, 2026 · 47 min

- Listen: [MP3, 45 MB](https://podcasts.robotpaper.ai/aws/ep/2026-10-30-week/2026-10-30-week-aws.mp3)
- Subscribe via RSS: https://podcasts.robotpaper.ai/aws/feed.xml (paste into any podcast app)
- Web page: https://podcasts.robotpaper.ai/aws/ep/2026-10-30-week/
- Chapters (JSON): https://podcasts.robotpaper.ai/aws/ep/2026-10-30-week/chapters.json
- Transcript (text): https://podcasts.robotpaper.ai/aws/ep/2026-10-30-week.txt
- Show: [AWS Drama](https://podcasts.robotpaper.ai/aws/) (RSS: https://podcasts.robotpaper.ai/aws/feed.xml)

## Chapters

- `00:00:00` Intro
- `00:00:10` Bedrock search filters
- `00:11:44` Aurora capacity
- `00:15:49` Replacement defaults
- `00:16:22` SQL Server updates
- `00:19:20` Image backfill
- `00:20:43` Cross-region identity
- `00:22:38` Neptune index rebuild
- `00:23:57` Grafana upgrade
- `00:25:00` EKS Gateway API
- `00:25:59` Overlapping networks
- `00:28:13` Persistent chat
- `00:29:54` Airflow in Mexico
- `00:31:27` Corretto quarterly
- `00:35:05` Hydro code review
- `00:45:54` Outro

## Show notes

### AI and machine learning (17)

- Web Search on Amazon Bedrock now supports domain filtering and user location controls
- AWS Key Management Service now supports ML-KEM for post-quantum key encapsulation
- Amazon SageMaker Unified Studio adds light mode support for IDC-based domains
- OpenAI GPT-6.1 Terra is now generally available on Amazon Bedrock
- Amazon Bedrock now offers Canada cross-Region inference for Claude Haiku 5.5
- SageMaker MLflow Apps now support MLflow 3.16
- AWS Agent Registry is now available in four additional AWS Regions
- Quick now supports smart approvals for agentic actions
- Amazon Bedrock AgentCore CLI version 1.0 is now available
- AWS Step Functions workflows now available as AI agent tools
- Amazon SageMaker AI now supports SOCI-indexed images to start training jobs in seconds
- Accelerate agentic AI with NVIDIA Nemotron models on Amazon EKS and SageMaker HyperPod EKS
- AWS named a Leader in the 2026 Gartner Magic Quadrant for Contact Center as a Service
- bedrock-agentcore: Adds AgentCore Runtime session endpoint support, including sessionEndpointUrl on the InvokeAgentRuntime response
- bedrock-agentcore-control: Adds Session Endpoint Feature on AgentCore Runtime
- bedrock-agentcore-control: Adds the ResetTokenVaultKeyMaterial API, which restores a token vault whose customer managed key was deleted by moving it to an AWS owned key.
- sagemaker: Amazon SageMaker HyperPod adds the MultiNodeTrainingValidation deep health check for clusters orchestrated by Amazon EKS.

### Serverless (6)

- Amazon Aurora serverless now supports up to 384 ACUs
- AWS IAM Identity Center's Identity Store now supports change notifications through Amazon EventBridge
- AWS Lambda durable functions now exports live telemetry for execution insights
- Introducing time series collections for the next generation of Amazon OpenSearch Serverless
- AWS Lambda supports Seekable OCI (SOCI) lazy loading for Lambda Managed Instances
- Public endpoints for Amazon ElastiCache Serverless now in AWS GovCloud (US)

### EC2 (1)

- AWS EC2 Auto Scaling now supports Launch Before Terminate at account-level to replace instances

### Databases (7)

- Amazon RDS now supports the latest Cumulative Update and GDR updates for Microsoft SQL Server
- Amazon ElastiCache for Valkey now supports changing durability on existing clusters
- Amazon ElastiCache expands data tiering to R8gd and R7gd nodes in Europe (Spain) and Asia Pacific (Hyderabad)
- Amazon RDS for Oracle now supports September 2026 Monthly Recommended Patch
- Amazon DynamoDB Accelerator (DAX) is now available in additional Regions
- Amazon Aurora now supports post-quantum TLS key exchange
- Amazon RDS for Oracle and SQL Server now support storage autoscaling on additional storage volumes

### Containers (3)

- Amazon ECR now supports replication backfill for existing container images
- Amazon ECS now supports GPU sharing with NVIDIA Multi-Process Service (MPS)
- Amazon ECS Managed Instances adds cross-instance and Spot consolidation

### Security Identity & Compliance (3)

- AWS IAM Identity Center now supports cross-Region trusted identity propagation
- Amazon EKS Dashboard now provides fleet-wide visibility into certificate authority (CA) rotation
- ACM now supports ACME for public certificates in AWS GovCloud (US)

### Storage (3)

- Amazon Neptune now supports online index rebuild to reclaim storage
- AWS Backup adds restore testing support for Amazon Redshift
- Build RAG-ready indexes from Amazon S3 with Amazon OpenSearch intelligent ingestion

### Management & Governance (2)

- Amazon Managed Grafana now supports in-place upgrade to Grafana version 13.2
- AWS Backup now supports cross-account management and cross-account backup in the AWS European Sovereign Cloud

### More core services (2)

- Amazon EKS Auto Mode now supports Kubernetes Gateway API
- AWS becomes first CSP to complete the Cloud High profile assessment with 172 services and features in scope

### Networking & Content Delivery (2)

- AWS Transit Gateway and AWS Cloud WAN now support 1:1 NAT
- Your AWS Networking and Content Delivery guide to re:Invent 2026

### Contact Center (5)

- Amazon Connect Customer Global Resiliency now supports persistent chat across linked AWS Regions
- Amazon Connect Customer now measures the outcomes that drive retention and revenue
- Amazon Connect Customer now lets you view metrics by customer lifetime value and repeat-caller status
- Amazon Connect Customer expands testing and simulation to validate flow action outcomes
- Amazon Connect Customer supports inline attachment previews

### Analytics (4)

- Amazon MWAA now available in Mexico (Central) Region
- Amazon EMR on EKS reduces EMR charge by 30 percent
- AWS Glue Data Catalog now supports OAuth for Apache Iceberg REST Catalog APIs
- Amazon MSK Connect now supports updating connector plugins and worker configurations

### Everything else (9)

- Amazon Corretto October 2026 Quarterly Updates
- Amazon Corretto October 2026 Patch Updates
- Amazon IVS Low-Latency Streaming now provides a failover slate
- Announcing the AWS Marketplace and Amazon Business unified storefront
- AWS Marketplace now enables partners to list once and sell through multiple subsidiary accounts
- Amazon WorkSpaces Core Managed Instances now supports ModifyVolume API
- AWS Elemental MediaConnect Router now supports endpoint connection management
- AWS DataSync Enhanced mode now supports Amazon FSx for Windows File Server, FSx for NetApp ONTAP, and Red Hat OpenShift Virtualization
- Recommended resilience testing in the next generation of AWS Resilience Hub

### Around the web (not official AWS news; past week) (12)

- AWS’s Hydro team wants the compiler to catch coding agents’ mistakes
- Andy Jassy’s prediction: Amazon could have two $1 trillion businesses within a decade
- AWS growth hits 39%, its fastest in nearly five years, as Amazon misses on earnings
- Show HN: Wraps – Resend-style email on your own AWS SES, open source
- MCP Token Cost: I Cut My MCP Server From 15 Tools to 4
- Never Put a Pet on an AWS Spot Instance
- AWS Elastic Beanstalk for regulated workloads: Built for the next decade, not just the first deploy
- 49, recently laid off Sys Admin, could use some encouragement!
- AWS Transit Gateway now supports 1:1 NAT (SNAT/DNAT) for connecting overlapping networks
- AWS Cost Updates - 2026-10-23
- AMA: Building on AWS with the coding agent you already use (10/29)
- Lambda static IP without a NAT Gateway

### Music

"Sneaky Snitch" Kevin MacLeod (incompetech.com)\
Licensed under Creative Commons: By Attribution 4.0\
http://creativecommons.org/licenses/by/4.0/

The voices you're hearing are AI-generated; the stories come from public sources linked in the show notes.\
Music: Eleven Music.\
An independent project; not affiliated with or endorsed by Amazon Web Services, Y Combinator, Hacker News or any vendor or news outlet mentioned. Opinions are the hosts' own.\
Produced by Roy Osherove.

## Transcript

\[Linus\] This is RobotPaper's weekly AWS news podcast; the voices you're hearing are AI-generated; the stories come from public sources.

\[Dennis\] Artificial intelligence and machine learning. This week's updates include controls over model inputs, agent actions, and the infrastructure that runs them. Amazon Bedrock is a managed service for generative artificial intelligence, with interfaces for running foundation models. Its Web Search tool now supports domain filtering and user location controls to make the search results grounding a response more specific.

The filters parameter accepts up to one hundred allowed domains to restrict results to chosen sources, or up to one hundred blocked domains to exclude unwanted sources. Separately, optional location parameters help search return results relevant to a particular geography, such as regional availability or local services. Both controls are configured on the existing Web Search tool in an OpenAI Responses A P I request.

The change is useful when a response needs evidence from a defined set of sites, or when the same question has different answers in different places. The controls are available wherever Web Search is supported: three commercial U S regions and GovCloud U S West.

\[Linus\] AWS Key Management Service protects cryptographic keys, including private keys held in hardware security modules. K M S now supports M L K E M, a post-quantum standard for establishing a shared secret between two parties.

The purpose is to address the harvest now, decrypt later threat: an adversary records a key exchange today, then tries to recover its secret when sufficiently powerful quantum computers become available. The launch supports all three key specifications in the standard. An encapsulation operation uses the public key to produce a random thirty-two-byte shared secret and its encrypted form. The corresponding private key recovers the secret through decapsulation.

That private key stays protected in hardware security modules validated at FIPS one forty dash three, level three. The public key can be exported to partners, client applications, or external systems so they can perform encapsulation locally with a cryptographic library. Support extends across K M S regions except the Middle East regions in the United Arab Emirates and Bahrain, and the China regions.

\[Ada\] Amazon SageMaker Unified Studio is a development environment with configurable domains and interface settings. It now supports light mode for domains based on AWS IAM Identity Center. Studio's appearance setting can select light, dark, or the device's own theme, persists across browsers and devices, and is available in all regions where Studio is offered.

\[Ken\] OpenAI G P T six point one Terra is generally available on Bedrock, adding a model for everyday workloads that combine text, images, and agent actions. It accepts text and image inputs with a context window of up to one million input tokens.

The announcement positions Terra for tasks such as drafting proposals, interpreting technical drawings, investigating codebases, and operating across business systems. It reports input and output token prices at least fifty percent lower than G P T six point one Sol and G P T five point six Terra.

In OpenAI's Agents' Last Exam evaluation, the new model scored fifteen percentage points above the previous Terra model and performed on par with G P T six Astra. The announcement says it completed tasks more than three times faster than Astra and six point one Sol in that evaluation. Access is through supported Bedrock interfaces, with ChatGPT Work and Codex also configurable to use the model on Bedrock.

\[Grace\] Bedrock now offers Canada cross-region inference for Anthropic's Claude Haiku five point five, routing requests only between Canada Central and Canada West in Calgary so inference processing stays in Canada. The profile has limited model selection, and future models or versions might not be released for it; existing U S and global inference options remain supported for workloads without that Canadian processing requirement.

\[Dennis\] SageMaker AI supports machine learning training and processing workflows; its MLflow Apps track experiments and help inspect generative application traces. MLflow Apps now support MLflow three point sixteen, with a redesigned trace explorer and faster charts for experiments with large metric histories. New apps use that version by default, and existing apps will upgrade during an upcoming maintenance window without customer action.

\[Linus\] AWS Agent Registry is a managed catalog for organizing and discovering agents, tools, skills, and other organizational resources, with search and an approval workflow. Agent Registry has expanded to London and Frankfurt, plus two additional regions, bringing availability to nine regions.

\[Ada\] Amazon Quick supports agents that take actions such as sending email, updating records, and creating tickets. Quick now adds smart approvals: a separate review model evaluates eligible actions against the user's request, automatically approving routine work while referring consequential actions to a person.

Existing permissions take precedence. The reviewer only considers actions configured to ask for approval each time; denied actions remain denied, and the feature never grants new access. Sending content outside the organization, changing permissions, deleting content, making payments, and going beyond the request all require human review. Each automatic approval applies only to the exact action evaluated, and a failed or timed-out review also goes to the user.

Smart approvals is enabled by default. Administrators control eligibility through custom permissions, while users can switch it off for conversations and scheduled runs. AWS CloudTrail provides the action-review record and records changes to the setting. The feature is available wherever Quick's agentic capabilities are supported.

\[Ken\] Bedrock AgentCore provides resources for developing and operating agents, including runtimes, memory, identity, and gateways. AgentCore's command-line interface version one point zero is generally available, adding consistent commands and direct work with resources without first placing them in a project. Existing projects can migrate with the migration command, and developers can invoke resources by identifier, open an interactive shell, and use guided resource selection.

\[Grace\] AWS Step Functions orchestrates defined, multi-step workflows; AgentCore Gateway connects agents to tools through the Model Context Protocol. Step Functions Standard and Express workflows can now become Gateway tools, allowing an agent to choose when to start a process while Step Functions follows its defined sequence, such as checking refund eligibility, obtaining approval, and issuing credit. Agents can also inspect workflow runs, and the integration is available in all Gateway regions at standard service rates.

\[Dennis\] SageMaker AI now supports Seekable O C I indexes for training and processing container images, allowing jobs to start before the entire image downloads. The service fetches the files needed to start the container and streams the remainder in the background as code runs.

Deep learning images routinely exceed ten gigabytes, and downloading them adds billable time before training begins. The company reports container starts in seconds and up to ninety percent less graphics processor idle time during image pulls in its benchmarks, with no impact on training performance.

Amazon Elastic Container Registry stores the private container images used here. Adoption means selecting a pre-indexed Deep Learning Container image or adding an index to an existing E C R image with open-source tools, without rebuilding it. Job definitions, entry points, and data channels stay the same. SageMaker AI detects the index automatically; if it cannot use one, the job continues with a standard download. Support is available across all SageMaker AI regions.

\[Linus\] The machine learning blog describes deploying NVIDIA Nemotron three point five Super V L on managed Kubernetes infrastructure. Amazon Elastic Kubernetes Service manages Kubernetes environments, while SageMaker HyperPod supports clusters used for machine learning workloads. The article covers E K S, its Auto Mode, and HyperPod clusters orchestrated by E K S.

Nemotron is an open multimodal model with one hundred twenty billion total parameters and twelve billion active per token. It accepts text, images, documents, and video, produces text, and supports contexts of up to one million tokens. NVIDIA publishes the weights, training datasets, and post-training recipes, allowing customization and operation on infrastructure the customer controls.

The source highlights document analysis involving tables and charts, coding workflows, web research, security analysis, and video incident analysis. Although only a subset of parameters activates for each token, graphics processor memory still needs to hold the full weights. The deployment guidance therefore distinguishes hardware and precision configurations, including requirements that vary substantially between the available checkpoints.

\[Ada\] Amazon Connect Customer provides cloud contact center capabilities, with customer context and agentic actions supporting service interactions. In its machine learning blog, the company reports that Gartner named it a Leader in the twenty twenty-six Magic Quadrant for Contact Center as a Service for Connect, for the fourth consecutive year. The post emphasizes personalized interactions, auditable actions, and human intervention as features of Connect Customer.

\[Ken\] Spotted in the S D K changelog: AgentCore Runtime adds session endpoint support, including a session endpoint U R L in the response from the runtime invocation operation.

\[Grace\] Also spotted in the S D K changelog, the AgentCore control interface adds the Runtime Session Endpoint feature. The entry supplies no further configuration details.

\[Dennis\] Spotted in the S D K changelog: AgentCore adds an operation to restore a token vault after its customer-managed encryption key was deleted, by moving the vault to a company-owned key. The reset permanently deletes stored OAuth two point zero tokens, requiring users to consent again; the same entry adds a client authentication override option for on-behalf-of token exchange.

\[Linus\] Spotted in the S D K changelog: HyperPod adds a multi-node training validation health check for E K S-orchestrated clusters. It runs a short training workload across NVIDIA graphics processor instances and flags instances that produce numerical errors or take longer than expected.

\[Ada\] Serverless. These updates address database capacity, event-driven automation, and the operational data around long-running applications. Amazon Aurora Serverless automatically adjusts database capacity with workload demand. Aurora Serverless now supports up to three hundred eighty-four capacity units for its PostgreSQL-compatible edition, a fifty percent increase that provides up to seven hundred sixty-eight gibibytes of memory. Existing clusters on platform version four can raise their maximum capacity setting, with an engine upgrade if needed, while supported configurations can still scale down to zero between bursts of work.

\[Ken\] Identity Center's Identity Store is its central repository for users, groups, and memberships, and Amazon EventBridge carries events that can trigger automated workflows. Identity Store now sends real-time change events to EventBridge when identities, memberships, or the store itself are created, updated, or deleted. The events can automate provisioning audits, security alerts, downstream access updates, and onboarding or offboarding tied to group membership.

\[Grace\] AWS Lambda runs application code without requiring customers to manage servers, and its durable functions support multi-step workflows lasting up to a year. Lambda's new Durable Execution Telemetry Exporter captures step histories and sends status, timing, inputs, outputs, and errors to thirteen built-in destinations, with sensitive-field redaction and sampling controls. A companion Model Context Protocol server lets agents query the exported data in natural language, supporting failure investigation and history retention beyond the standard period.

\[Dennis\] Amazon OpenSearch Serverless ingests, stores, and analyzes data without customers provisioning or managing clusters. Its next-generation offering now supports time series collections for logs, metrics, traces, and event data, with hot and warm storage in the same collection.

The new warm tier is intended for older, less frequently accessed data, and customers choose how long information remains hot before it moves. Each collection supports up to one hundred terabytes of hot storage. The service automatically absorbs ingestion spikes, while retaining searchable history over months or years.

The launch also adds alerting and anomaly detection, allowing teams to identify unusual patterns in log and event data and receive notifications without separate tooling. The source identifies application monitoring, security analytics, and connected-device telemetry as target workloads. The combination addresses a recurring observability tradeoff: keeping recent events fast to query while retaining older records at lower storage cost. Time series collections are available in every region where OpenSearch Serverless is offered.

\[Linus\] Lambda also adds Seekable O C I lazy loading for Event Functions packaged as container images on Lambda Managed Instances. It fetches files as needed during execution environment startup, reporting startup reductions of up to seventy percent for workloads such as data processing, high-throughput application backends, and batch computation; adoption requires an image index stored with the image in E C R.

\[Ada\] Amazon ElastiCache Serverless provides a managed cache without infrastructure provisioning. Its Valkey public endpoints now reach GovCloud U S, allowing applications outside a virtual private cloud to connect without a private network tunnel, with identity-based authentication over T L S one point three. Public endpoints require Valkey nine point zero or later, add no charge beyond standard pricing, and are now available across commercial, GovCloud, and China regions.

\[Ken\] Core services. Compute. Amazon EC2 Auto Scaling adjusts groups of virtual machines and replaces unhealthy or aging instances. Launch Before Terminate can now be the account-wide default, so replacement instances must pass health checks before the old instances terminate during refreshes, rebalancing, health-related replacements, and maximum-lifetime replacements. The setting applies across Auto Scaling groups and is available in all regions at no additional cost.

\[Grace\] Databases. This week's changes cover database maintenance, durability, and capacity management. Amazon Relational Database Service, or RDS, manages relational database instances. RDS for SQL Server now supports updated cumulative and general distribution releases for SQL Server twenty seventeen, twenty nineteen, twenty twenty-two, and twenty twenty-five; the announcement recommends upgrading existing instances to apply the fixes.

\[Dennis\] ElastiCache for Valkey can now enable or disable durability on existing node-based clusters, allowing an existing cache to become a persistent data store without replacing the cluster. Synchronous writes persist across at least two availability zones before acknowledging the client, with single-digit millisecond write latency; asynchronous writes retain microsecond write latency but can lose up to ten seconds of uncommitted data in a failure.

\[Linus\] ElastiCache data tiering now supports R eight G D nodes in Spain, and R seven G D nodes in Spain and Hyderabad. Tiering moves less recently used data from memory to local solid-state storage, fitting workloads that regularly access up to twenty percent of their dataset and tolerate extra latency when fetching data from disk.

\[Ada\] RDS for Oracle now supports the September twenty twenty-six Monthly Recommended Patch for Oracle Database nineteen C and twenty-six A I, combining the September critical security patch with additional fixes on top of July's release update. It also includes time zone file version forty-six, covering rule changes in parts of Canada and North Africa that matter for stored timestamps with time zone information.

\[Ken\] Amazon DynamoDB Accelerator, or DAX, is a managed in-memory cache for Amazon DynamoDB that accelerates database reads. DAX is now available in Canada Central and Cape Town, allowing read-intensive applications such as gaming leaderboards and retail catalogs to keep their cached data within those regions.

\[Grace\] Aurora PostgreSQL now supports post-quantum T L S key exchange for protecting data in transit. PostgreSQL-compatible versions eighteen and higher allow administrators to modify the S S L groups parameter and choose cryptographic groups from the supported allow list.

\[Dennis\] RDS for Oracle and SQL Server now support storage autoscaling on additional storage volumes, extending the automatic capacity growth previously available on the primary volume. Administrators set a maximum for each additional volume, with support for new and existing instances, zero downtime during scaling, and no autoscaling charge beyond the provisioned storage.

\[Linus\] Containers. These updates address existing image catalogs and how container tasks use compute capacity. E C R replication backfill can now copy images that existed before replication was configured, using a single A P I call instead of manual retagging; teams can monitor per-image results, run up to ten backfill jobs concurrently, and stop a job while retaining images already copied.

\[Ada\] Amazon Elastic Container Service, or ECS, runs and schedules container tasks. It now supports sharing a single NVIDIA G P U across multiple tasks through NVIDIA Multi-Process Service, with per-container memory requirements and optional compute-share limits. Supported on Managed Instances and EC2, the capability targets workloads that use only part of a G P U, including small-model inference, embedding generation, and multi-model serving.

\[Ken\] ECS Managed Instances provisions, configures, and operates EC2 instances for container workloads. It now evaluates tasks across multiple underutilized instances together to consolidate them onto fewer instances, and extends consolidation to Spot capacity after checking that replacement capacity is available.

\[Grace\] Security, identity, and compliance. Identity Center now carries trusted user identity across regions. Trusted identity propagation passes the signed-in user's identity and group associations along with a request, so downstream services can authorize the individual and record who accessed a resource. Previously, all services using that identity context had to operate in the same region. Reaching another region meant using broader shared roles or bringing the services together in one region. The new support removes that regional restriction while preserving permissions tied to the individual user. That matters when compute runs separately from the data it queries, when workloads move to regions with available capacity, or when an architecture spans regions for resilience. The accompanying technical blog explains that identity objects, including users, groups, and session information, are replicated to the regions an organization chooses. Cross-region trusted identity propagation is available wherever Identity Center is offered, at no additional cost.

\[Dennis\] The EKS Dashboard now shows certificate authority rotation status across clusters, including approaching expirations and upcoming automated rotation events. Clusters created since twenty eighteen have authorities with ten-year validity periods, so the consolidated view helps platform teams prioritize rotation work across their accounts and regions.

\[Linus\] AWS Certificate Manager, or ACM, issues and manages T L S certificates. Its ACME protocol support now extends to both U S GovCloud regions, providing a managed endpoint for public certificate issuance and renewal through compatible clients such as Certbot and cert-manager.

\[Ada\] Storage. Amazon Neptune Database runs Gremlin, openCypher, and S P A R Q L query workloads. Neptune now supports rebuilding indexes online to reclaim space left by inserts, updates, and deletions while read and write queries continue on the writer and replicas. Administrators can target all indexes or specific index types, with concurrency and write-buffer controls to balance rebuild speed against workload impact.

\[Ken\] AWS Backup manages backups and recovery testing across supported resources. Amazon Redshift runs analytics queries. Backup now supports scheduled restore testing for Redshift, automatically selecting recovery points and measuring completion time against recovery objectives; optional validation can check the restored resource before the test resources are deleted.

\[Grace\] Amazon S3 stores objects, including documents and structured data. A new post on the company’s Big Data Blog introduces OpenSearch intelligent ingestion for building retrieval-ready indexes from S3, describing the work involved in parsing document layouts and tables, interpreting visuals, and creating embeddings for search.

\[Dennis\] Management and governance. Amazon Managed Grafana provides managed workspaces for Grafana dashboards. In-place upgrades from version twelve point four to thirteen point two are now supported, adding Git Sync for version-controlled dashboards and dynamic dashboards whose layout responds to data and variables. Amazon CloudWatch ingests metrics through an OpenTelemetry endpoint; the release adds Prometheus Query Language support for querying those metrics in its Grafana data source plugin.

\[Linus\] Backup now supports cross-account management, cross-account copies, and delegated administrators in the European Sovereign Cloud's Germany region. Organizations can apply shared backup policies, monitor jobs centrally, and copy recovery points into another account for protection against deletion in the source account, with standard Backup pricing and no additional charge for these capabilities.

\[Ada\] More core services. EKS Auto Mode now supports the Kubernetes Gateway A P I, provisioning application load-balancing infrastructure from Gateway and H T T P Route resources without a separately installed controller. It supports H T T P and G R P C routing, weighted traffic splitting, cross-namespace routing, and automatic certificate discovery through ACM, while existing Ingress resources continue to work alongside it.

\[Ken\] The company's Security Blog reports completion of Canada's Cloud High security assessment, covering one hundred seventy-two services and features, effective September eighteenth, twenty twenty-six. The announcement describes Amazon's cloud as the first provider to complete this assessment, which addresses integrity and availability for government and service-provider systems handling sensitive information or supporting services at national scale.

\[Grace\] Networking and content delivery. AWS Transit Gateway connects private networks through a central hub. AWS Cloud WAN builds and manages wide-area networks. Both now support managed one-to-one network address translation on their attachments, allowing networks with overlapping private address ranges to communicate without renumbering. The integration uses Regional Private NAT Gateway, with rules specifying which private range each source range translates into. Translation happens automatically in both directions, so there is no separate return-traffic rule or translation appliance to operate. This addresses a practical integration problem when companies merge, connect partner networks, or join independently provisioned virtual private clouds. The announcement also identifies multicloud connectivity as a use case. The change is about preserving existing address plans while connecting networks that previously conflicted, with the translation capability operated and scaled by the provider.

\[Dennis\] The Networking and Content Delivery Blog has published its guide to re:Invent, running November thirtieth through December fourth in Las Vegas. The networking track includes advanced virtual private cloud design, private service connectivity across accounts and organizations, D N S developments, load balancing, and connectivity across cloud partitions. The guide distinguishes the formats: lecture-style breakouts, focused lightning talks, interactive chalk talks, coding demonstrations, hands-on workshops, and small-group builders' sessions. Lightning talks are walk-up only; other formats accept reservations through the event portal or walk-ups when space is available. Related Cloud Operations posts cover agent-assisted incident investigation, remediation controls and auditability, and observing A I agents in production. These published guides provide session themes and planning information ahead of the event. The networking post also notes that scheduled dates and venues can change, so its timetable remains subject to updates.

\[Linus\] Contact Center. This week's Connect Customer updates cover conversation continuity, outcome measurement, and agent workflows. Global Resiliency now carries persistent chats across linked regions, including conversation context, metadata, transcripts, and attachments; adoption requires onboarding to Global routing.

\[Ada\] Connect Customer now scores whether interactions ended in satisfaction, a purchase, a risk of churn, or successful self-service. Businesses can customize the default definitions and access these outcome metrics through analytics dashboards and A P Is wherever the service is offered.

\[Ken\] Connect Customer metrics can now be broken down by customer lifetime value and repeat-caller status. These views help identify problems hidden in aggregate results, such as repeat callers struggling with self-service or satisfaction falling among high-value customers.

\[Grace\] Connect Customer testing and simulation can now validate the data a flow action returns when it completes, beyond checking that the action started. Tests can inspect a Lambda response or variables from an Agentic C X block and identify the specific action that produced an unexpected result.

\[Dennis\] Connect Customer agents can now preview P D Fs, images, and text files inside their workspace across email, chat, and cases, including multipage documents. This limits the need to download sensitive customer files and is available in North Virginia, Oregon, and eight more regions worldwide.

\[Linus\] Analytics. Amazon Managed Workflows for Apache Airflow, or M W A A, manages orchestration for end-to-end data pipelines. Its provisioned environments are now available in Mexico Central, where the serverless option was already offered, allowing provisioned workflows to run closer to local data sources and end users.

\[Ada\] Amazon E M R on E K S runs Apache Spark workloads on E K S. Its added compute and memory charges have fallen by thirty percent across all supported regions, automatically for existing workloads; the underlying E K S charges are unchanged.

\[Ken\] AWS Glue Data Catalog provides a catalog that analytics engines can use to access Iceberg tables. Its Iceberg REST Catalog A P Is now accept OAuth two point zero bearer tokens, allowing partners to exchange an identity token for a short-lived, Glue-scoped token without receiving cloud credentials; existing access permissions still govern the tables.

\[Grace\] Amazon Managed Streaming for Apache Kafka Connect, or M S K Connect, deploys, monitors, and scales Kafka Connect connectors. Existing connectors can now update their custom plugins and worker configurations in place, keeping committed offsets so processing continues from where it stopped after the update, without the previous delete-and-recreate operation or an additional charge.

\[Dennis\] Everything else. Amazon Corretto is a no-cost, multiplatform distribution of Open J D K for production Java applications. Its October quarterly release supplies security and critical updates across the Java twenty-seven, twenty-five, twenty-one, seventeen, eleven, and eight lines, available through downloads and the Linux package repositories.

\[Linus\] A subsequent October patch update is also available for Corretto across those Java release lines. The October twenty-six announcement lists newer builds for Java twenty-seven through eleven, while retaining eight update five hundred twelve for Java eight, with downloads and package-repository installation available.

\[Ada\] Amazon Interactive Video Service, or I V S, provides managed low-latency and real-time video streaming. Low-Latency Streaming channels can now display an animated failover slate during an unexpected interruption, then resume playback automatically when the broadcaster reconnects; the feature supports new and existing channels.

\[Ken\] AWS Marketplace offers software, data, and A I tools for purchase. A unified storefront now combines those listings with physical goods from Amazon Business under shared governance rules: Marketplace purchases finish inside the storefront, while Business purchases transfer through single sign-on to an authenticated cart.

\[Grace\] Marketplace partners with multiple legal entities can now publish a listing once and authorize subsidiaries to sell it through direct private offers under their own local tax identities. The primary account controls eligible products and buyer countries and receives consolidated reporting; revoking authorization stops new offers while leaving existing offers and agreements unchanged.

\[Dennis\] Amazon WorkSpaces Core Managed Instances manages instances and their storage through an A P I and partner solutions. Its new Modify Volume operation changes volume size, supported type, input-output performance, or throughput in place, without detaching storage or restarting the instance, replacing the previous requirement to rebuild the managed instance.

\[Linus\] AWS Elemental MediaConnect Router routes video between inputs and outputs for media workflows. Its new endpoint registry connects external devices or software and shares those endpoints across accounts, reducing manual coordination of addresses and encryption settings; the registry is available without an additional charge wherever the router is supported.

\[Ada\] AWS DataSync transfers data between storage locations. Enhanced mode now supports Amazon F S x for Windows File Server and F S x for NetApp ONTAP, file-storage destinations that can also serve as sources, and its agents can run on Red Hat OpenShift Virtualization version four point twenty or later; Enhanced mode now has the same location options as Basic mode.

\[Ken\] AWS Resilience Hub helps teams assess application resilience against defined policies. An architecture blog describes recommended resilience testing, which uses known resources, dependencies, and policies to create fault-injection experiments, then evaluates recovery against success criteria and the recovery time objective to produce a pass or fail result.

\[Grace\] Around the web and R slash AWS. The New Stack reports on Hydro, the open-source Rust framework for distributed systems, and its team's effort to make the compiler catch mistakes made by coding agents. Agents now write most of the framework with humans directing them. A recent change checked whether a distributed system remains correct when it resends messages after a failure, and produced a pull request of roughly five thousand lines. Its interdependent components made it difficult to split into smaller reviews. Research lead Shadaj Laddad described the problem at AGNTCon and M C P Con in San Jose: agents let the team attempt more, but the resulting changes arrive faster and grow longer. A feature can work by itself and then violate guarantees when combined with another feature. Laddad also questioned relying solely on unit tests when the same agent writes both the implementation and those tests. The article's central concern is how correctness checks can keep pace with agent-generated changes in software spread across multiple machines.

\[Dennis\] GeekWire reports that Andy Jassy believes both Amazon's cloud business and its stores business could each reach a trillion dollars in annual revenue within five to ten years. His explanation for the cloud forecast centers on contracted capacity and growing A I use. The cloud business has an annual revenue run rate of one hundred eighty-four billion dollars, and Jassy pointed to five hundred seventy-eight billion dollars in contracted future revenue. He said large customers sometimes reserve capacity in blocks as large as a gigawatt, providing visibility into demand through twenty twenty-seven, twenty twenty-eight, and part of twenty twenty-nine. He also expects new applications to emerge as model-running costs fall, citing models finding and fixing security flaws. Jassy acknowledged uncertainty in projecting so far ahead at those growth rates. GeekWire places the forecast in context: in March he had discussed a cloud business reaching at least six hundred billion dollars in annual revenue, a substantially lower potential milestone.

\[Linus\] In separate earnings coverage, GeekWire reports that cloud revenue grew thirty-nine percent to forty-six billion dollars in the third quarter, its fastest growth in nearly five years. The broader company's results show why that growth did not settle investor concerns. Earnings per share were a dollar eighty-nine, below the dollar ninety-seven analysts expected, with higher interest costs and other losses weighing on profit. Operating income nevertheless reached twenty-eight point one billion dollars, above expectations, and the operating margin was thirteen point nine percent, the highest in Amazon's history. Shares fell more than seven percent after hours following a four percent decline during the regular session. Spending remains part of the story: Jassy said capital investment in twenty twenty-seven would be meaningfully higher than the roughly two hundred twenty billion dollars planned for this year. The report connects strong cloud demand with the financial burden of expanding capacity, as investors weigh growth against the company's rising investment and financing costs.

\[Ada\] On Hacker News, the open-source project Wraps presents an email platform that deploys into a customer's own cloud account. Amazon Simple Email Service, or S E S, handles email sending, while Wraps supplies a dashboard for sending messages, running broadcasts, and building workflows, plus a typed TypeScript S D K that sends through S E S. The project says delivery, open, and click events go into a table the customer owns. Broadcast features include scheduling, contact segments, double opt-in, and one-click unsubscribe. Automation can respond to a signup, an event, a topic change, or a schedule, then wait or branch according to what a contact does. Wraps also monitors bounce and complaint rates and domain authentication, aiming to surface sending problems early. Its stated commercial model separates the platform charge from sending charges paid directly to the cloud provider. The project page positions this as a way for teams to get managed campaign and workflow controls while keeping the underlying sending infrastructure in their own account.

\[Ken\] On dev dot to, Sarvar Nadaf describes rebuilding a read-only cloud cost-triage M C P server in the article M C P Token Cost: I Cut My M C P Server From Fifteen Tools to Four. The author connected the server to a coding agent and asked where money was being wasted in an account. The original design needed seven tool calls to answer. After the redesign, the same account and question required two calls, and the tool schemas were eighty-eight percent smaller. Nadaf attributes the change to following the company's tool-design guidance and consolidating the available tools. The useful issue here is the context consumed by tool definitions before an agent begins its actual investigation, alongside the number of calls needed to complete a task. Both server versions are published together in the author's cloud cost-triage repository, allowing developers to compare the implementations. The post offers a concrete account of how tool interface design changed the cost-investigation workflow, with the before-and-after measurements supplied by its author.

\[Grace\] Another dev dot to post about selecting Spot instances describes a migration decision that turned on where a server kept its data. Author Maame Afua Fordjour was moving a server to another region and looking for the lowest cost. The attractive option was Spot capacity, which uses spare EC2 capacity at a discount. A teammate then asked where the server stored its state. In this case, its context lived on its own disk, with no separate database or storage holding a copy. The author explains that Spot capacity can be reclaimed, with interrupted instances terminated by default, and describes the short warning before interruption. That made the workload's dependence on one machine central to the decision. The post is a practical account of how a low compute price can obscure an application's recovery requirements: this server could not simply be discarded and recreated with its state intact. The useful experience is the storage question that changed the migration plan before the move.

\[Dennis\] The New Stack's article on AWS Elastic Beanstalk for regulated workloads argues that platform selection needs to account for years of operation after the initial deployment. Elastic Beanstalk is presented as a managed application platform in that discussion. The article focuses on the operational demands of long-running systems: operating systems need patches, runtimes reach the end of support, security requirements change, teams change, and auditors need evidence. It uses patient scheduling, financial applications, government services, and aerospace workloads as examples of software that may remain in production for years. The author's framework weighs developer simplicity alongside compliance posture and service continuity. The practical question raised is how a team will keep an application supported, observable, auditable, and operational as those conditions evolve. The article is useful as a statement of platform-selection criteria for regulated organizations, with attention to the life of an application beyond how quickly its code can first be deployed.

\[Linus\] On R slash AWS, a discussion of Transit Gateway's managed one-to-one network address translation highlights a distinction in the new feature's description. The post points out that the announcement names Regional Private NAT Gateway as the integration behind translation on Transit Gateway and Cloud WAN attachments. The author says customers cannot currently create Regional Private NAT Gateways directly: private NAT gateways remain zonal, while regional public NAT gateways are available. That is a useful operational detail alongside the earlier launch coverage. The new attachment feature is described as managed translation for networks with overlapping address ranges, while the post is asking about a separately creatable private gateway resource. The useful part of the discussion is the resource-availability question that teams may encounter when translating the launch description into a network design.

\[Ada\] An R slash AWS ask-me-anything session on October twenty-nine focused on building cloud applications with existing coding agents. The organizers described recurring failures such as an agent using an A P I as it existed in its training data, or missing a permission needed for one service to invoke another. Participants included engineers and developer advocates working on the Agent Toolkit for AWS, which combines a managed M C P server with skills and plugins. The announced discussion covered building, deploying, and debugging; steering agents when they make cloud-specific mistakes; and granting access to real accounts with limited permissions and audit trails. Token and context costs were also on the agenda, along with how the toolkit compares with the separate AWS Labs M C P servers. The session was open to people using other agent tools as well. Its practical focus was the gap between generating plausible application code and configuring the service behavior, access, and operational details required to run that code.

\[Ken\] A link shared on R slash AWS leads to OutboundGateway's article Lambda Static I P Without a NAT Gateway. It addresses functions that must call a partner A P I from an allowlisted address. The vendor describes its service as an alternative outbound path that gives a function a pair of fixed I P addresses, and says its setup uses two environment variables in the function runtime. The article compares that approach with configuring a virtual private cloud, subnets, routing, and NAT gateways for fixed-address outbound traffic. This is a vendor's account of its own service, with the useful distinction being where that outbound network setup is managed. Its example focuses on a small backend function, such as handling a form submission for a static website, where the author argues that the surrounding network configuration can become a significant part of the deployment work.

\[Grace\] That concludes RobotPaper's weekly AWS news podcast. Across this week's coverage, the recurring themes were more control over agent actions, less infrastructure work around training and applications, and clearer operational visibility. The database and networking changes addressed capacity, durability, and connections across existing environments. Around the web, the discussion turned to correctness, state, and the practical work of operating what agents build. The full write-up is at robotpaper dot A I. That is the weekly report. Goodbye.

---

The voices you're hearing are AI-generated; the stories come from public sources linked in the show notes. An independent project; not affiliated with or endorsed by Amazon Web Services, Y Combinator, Hacker News or any vendor or news outlet mentioned. Opinions are the hosts' own. Produced by Roy Osherove. Privacy: We count visits to these pages with Cloudflare Web Analytics (cookieless). Google Analytics (which sets cookies): visitors in the EU/EEA, the UK and Switzerland are asked first and it runs only if they accept (we go by the country Cloudflare reports for the connection; if it is unknown, we ask). Elsewhere it loads by default. Anyone can turn it off with the Privacy settings button at the bottom of every page; the choice is kept for a year in this browser (local storage "analytics-consent"). We count feed fetches and downloads in aggregate; IP addresses are hashed and never stored. If you send feedback, we keep what you write (and your email if you give it) to read and reply. © 2026 Roy Osherove.
