The expense of technology transitions often stems not from the technology itself, but from the migration process. Transitioning a Hive table to Apache Iceberg previously required weeks of engineering effort. This week, Amazon S3 Tables condensed that workflow into three discrete steps: initiate, synchronize as frequently as needed, and finalize. The original data source remains available for writes except during a brief window at transition time. Concurrently, Amazon DynamoDB introduced the capability to ingest data from S3 into pre-existing tables.
A common thread across these announcements is that neither service requires halting existing operations. This principle extends to the agent-related developments unveiled this week. The refreshed Amazon Bedrock AgentCore Runtime appears as a single configuration parameter on your current runtime, while Web Answer functions as a tool integrated into a gateway you already operate. Here's what's new in AWS this week.
Key Announcements
AgentCore Runtime V2 brings elastic scaling and faster cold starts
The updated AgentCore Runtime, now generally available, addresses performance and cost through two primary improvements. Memory allocation now scales dynamically: sessions initialize with minimal memory, expand on demand, and release unused capacity rather than reserving peak allocation, meaning you only pay for actual consumption. Cold start performance improves through environment snapshots: AgentCore prepares your environment once, creates a snapshot, and restores that snapshot for each new instance. Testing showed P75 cold starts remained around 2 seconds for container images ranging from 200 MB to 2 GB, whereas the original runtime escalated from approximately 5.4 seconds to nearly 30 seconds as image size increased. Per-unit pricing is higher, but substantially lower GB-hour consumption typically results in reduced overall costs, though outcomes vary by workload.
To enable V2, set platformVersion to V2 during runtime creation or updates. Both operations require several minutes while snapshots are prepared, and your container must report health status on /ping within 120 seconds—ensure health checks only trigger after initialization completes. Documentation provides optimization guidance for V2 compatibility. The runtime is accessible across 10 Regions, with committed baseline pricing launching soon.
S3 Tables simplifies Hive-to-Iceberg conversion
Parquet tables registered in the AWS Glue Data Catalog can now migrate into S3 Tables without constructing custom pipelines or allocating compute resources. S3 Tables examines the source and generates a compatibility assessment for all objects in scope. Incremental synchronization copies only modified data, allowing production systems to continue accepting writes until the final cutover moment. Post-migration, you can query the table using Amazon Athena, Amazon SageMaker Unified Studio, or any Iceberg REST API-compatible engine. Pricing covers only transferred data, with no conversion fees, and objects in S3 Intelligent-Tiering retain their current tier assignment. At launch, the service operates within the same account and Region across 10 Regions.
DynamoDB bulk loading from S3
Amazon DynamoDB now accepts line-delimited DynamoDB JSON or Amazon Ion files from Amazon S3 for bulk insertion into existing tables through managed, asynchronous operations. Individual records can include their own condition expressions, enabling a single file to contain unconditional writes, insert-only operations, and version-controlled updates. Write operations consume dedicated system capacity and cost less than equivalent per-request pricing. An Accelerated tier addresses time-sensitive workloads, with documented performance showing sub-one-hour completion for 100 million 1 KB items. Important caveat: job cancellation does not roll back items already persisted.
Web Answer tool for AgentCore
Agents can now invoke Web Answer, a single tool that delivers synthesized, attributed responses to natural-language queries, eliminating the need to build custom search, retrieval, and citation systems. Web Answer independently plans searches, executes multiple queries in parallel, and determines when sufficient evidence exists. Integration occurs through attachment as a target on AgentCore Gateway and invocation via MCP, with optional restrictions using domain inclusion and exclusion lists plus publish-date filtering. The service launches in US East (N. Virginia). Documentation clarifies when Web Answer suits your needs versus Web Search, which returns ranked results for your model to process.
Lambda queue support for Apache Kafka
AWS Lambda now processes records from individual Kafka partitions across multiple parallel invocations. Traditional consumer groups restrict parallelism to partition count. Queue mode leverages Kafka share groups, requiring Kafka 4.2 or later with share groups enabled, and works with self-managed clusters and managed services like Confluent Cloud. The tradeoff: you forfeit ordering guarantees but gain at-least-once delivery semantics, requiring idempotent function logic. The Compute blog includes detailed guidance on poller scaling and stream mode retention decisions.
AWS Interconnect preview with Equinix
AWS Interconnect enables direct, managed connections from Equinix's 282-datacenter network to AWS without manual cross-connect provisioning or router configuration. Connectivity spans four redundant links across two locations with quad redundancy, automatic BGP and VLAN setup, a 99.99% availability guarantee, and MACsec encryption by default. Configuration requires selecting a Region, bandwidth tier, and AWS Direct Connect gateway, then entering the activation key into Equinix Fabric. The preview currently covers the San Jose metro and, as of last Thursday, Northern Virginia, offering 1 Gbps connections at no cost during the preview period.
Security Hub SBOM inventory
AWS Security Hub now displays a comprehensive software bill of materials across your organization's Amazon EC2 instances, Amazon ECR images, and Lambda functions, showing each package version and resource count. This inventory proves invaluable when widely deployed libraries face vulnerability disclosures. The feature builds on SBOMs already generated by Amazon Inspector, populating automatically wherever Inspector operates. Security Hub Essentials includes this capability at no additional charge.
Additional AWS Developments
Julian Wood's quarterly serverless recap covers Q3 2026 highlights ahead of re:Invent, including Lambda SnapStart for container images, 90-minute timeout support on Lambda Managed Instances, Lambda MicroVMs across 10 Regions, 1 MiB payload capacity in Amazon SNS, and the new AgentCore Runtime.
AWS contributed a binary log optimization to MySQL addressing large-transaction performance. When binlog is active, massive transactions hold a global commit lock during event copying, blocking all other commits. The optimization, running in Amazon Aurora MySQL since 2019, promotes the spill file into the binlog instead of duplicating it. AWS testing demonstrated commit latency for a 50 GiB transaction dropping from roughly 7 minutes to 2 to 2.5 seconds. The enhancement has shipped to community MySQL.
Cognizant's OrderServ food-ordering platform achieved up to 64% cost reductions through multi-tenant ECS consolidation. The platform migrated from individual Amazon ECS clusters per brand on AWS Fargate to a unified cluster on EC2 Auto Scaling groups with warm pools. The 64% savings figure applies to three-year Compute Savings Plans (47% with one-year plans), while scaling latency decreased from 5–7 minutes to 30–60 seconds.
Upcoming Events
- AWS re:Invent returns to Las Vegas November 30 through December 4, with event catalog and reserved seating now available for registration
- AWS Community Days in Belém, Brazil (November 14), JAWS FESTA 2026 in Akita, Japan (November 17), and Port Harcourt, Nigeria (December 12)
- AWS Builder Center offers networking opportunities, solution sharing, and development resources, with upcoming in-person and virtual events listed on the events page