Bedrock AgentCore Adds Persistent Runtime Instances
AWS introduced runtime instances, a new compute option within Amazon Bedrock AgentCore that complements the existing microVM runtime. Where AgentCore's microVMs cap out at 8-hour sessions, runtime instances run on AWS-managed EC2 infrastructure and can persist for up to 14 days, with session hibernation and resumption available to control cost during idle periods.
The feature targets production agent workloads that need extended, stateful execution: multiple agents can share a filesystem on the same host and invoke each other as tools within a shared session, enabling multi-agent collaboration patterns — AWS demoed a writer agent generating code and a reviewer agent analyzing it, both persisting state through a shared session store. Runtime instances support GPU acceleration, containerized deployments alongside zip-based packages, and work with any foundation model, not just Bedrock-hosted ones. Popular frameworks including CrewAI, LangGraph, LlamaIndex, and Strands are supported out of the box.
Getting started requires minimal packaging — an @app.entrypoint decorator plus a zip file or container image — and the feature reuses existing AgentCore identity controls, observability, and APIs, so teams already on AgentCore microVMs shouldn't need a new mental model to adopt it. Runtime instances support Linux on both ARM64 and x86_64, Python 3.11–3.14, and are available in six regions at launch: US East (Ohio, N. Virginia), US West (Oregon), and select Asia Pacific and Europe regions. Pricing follows standard EC2 rates plus an AgentCore orchestration management fee.
Read more — AWS News Blog
Amazon DynamoDB Adds Native Vector Search at GA
Amazon DynamoDB now supports vector search natively, reaching general availability without requiring a separate vector database alongside it. AWS reports single-digit millisecond query latency at 99%+ recall, with support for embeddings up to 4,096 dimensions and no practical storage limits thanks to DynamoDB's existing horizontal scaling model — the company says the feature is designed to handle trillions of vectors.
The new SearchVectors API supports three distance functions — Euclidean, cosine, and dot product — plus inline filtering on non-vector attributes, so a query can combine semantic similarity with structured conditions in a single call. Each query returns up to 100 ranked results, each carrying both a similarity score and the item's other operational attributes, which avoids a second round-trip to fetch metadata after a vector match.
For teams currently running a dedicated vector store alongside DynamoDB for RAG or semantic search use cases, this removes the need to keep two data stores in sync — embeddings are stored using the existing PutItem API as float lists, no new table type or migration required. The feature is fully serverless with pay-per-request pricing and automatic scaling, and it integrates with embedding models from Amazon Bedrock, Cohere, and OpenAI. It's available across all commercial AWS regions, including GovCloud.
Read more — AWS News Blog
AWS Weekly Roundup: Bedrock GPT Price Cuts and Managed Prometheus Collectors
AWS's August 3 weekly roundup led with a pricing change: Amazon Bedrock cut prices for OpenAI's GPT-5.6 models effective July 30, with the Luna variant seeing an 80% reduction to $0.20 per million input tokens and $1.20 per million output tokens, and the Terra variant getting a 20% cut. Both reductions apply automatically to existing Bedrock usage with no customer action required.
Amazon CloudWatch introduced fully managed collectors for Prometheus metrics, removing the operational burden of deploying and maintaining separate collector agents. The managed collectors support monitoring across Amazon EKS, EC2, ECS, MSK, and OpenSearch Service, consolidating infrastructure metrics collection under a single managed service rather than requiring teams to run and patch their own Prometheus exporters.
Other notable items from the roundup include general availability of AWS Interconnect for Oracle Cloud, enabling private multicloud connectivity between AWS and OCI without traversing the public internet; multi-region replication support for IAM Identity Center directories, improving disaster-recovery posture for centralized identity management; and Variant data type support in Amazon S3 Tables' Apache Iceberg v3 integration, aimed at semi-structured data like IoT telemetry and application logs.
Read more — AWS News Blog
Google Cloud Cortex Framework v7 Reaches General Availability
Google Cloud released version 7 of Cortex Framework, its data integration platform for turning SAP ERP and SAP Business Data Cloud records into AI-ready data products in BigQuery. The core pitch is letting enterprises deploy AI agents against SAP data without touching mission-critical ERP systems directly — Cortex Framework translates raw SAP tables, including SAP-specific logic like TCURX currency decimal shifts and custom Z-fields, into semantically labeled data products that Gemini-based agents can reason over.
The headline new capability is an agentic data product builder: using natural-language prompts through Google Antigravity and Gemini, teams can generate custom data products without hand-writing transformation logic. Pipelines are now built on Dataform, giving version-controlled SQL, automatic dependency-graph resolution, and support for compiling and deploying against SAP ECC, S/4HANA, and SAP BDC in parallel from a single framework.
Cortex Framework v7 also ships an extensibility guide that separates framework-managed content from custom models, so teams can pull in framework updates without clobbering their own customizations, plus new solution samples for common use cases like sales pipeline health and procurement inefficiency detection. Incremental loading is the default processing mode, and organizations with existing CDC pipelines can plug those in directly rather than relying on Cortex's built-in change-data-capture handling. Schlumberger (SLB) is cited as an early validation customer, calling out the agentic data product builder as accelerating data-product creation for its own agentic AI initiatives.
Read more — Google Cloud Blog
Links & Sources
- Runtime instances: persistent compute for production AI agents on Amazon Bedrock AgentCore
- Amazon DynamoDB now supports real-time vector search at any scale
- AWS Weekly Roundup: Price reduction of GPT models in Bedrock, CloudWatch managed collectors for Prometheus metrics, and more
- Google Cloud Cortex Framework version 7 is now generally available
You May Also Enjoy
