AWS AI Updates: August 7, 2026

1. Amazon Bedrock AgentCore runtime instances reach general availability

AWS. Amazon Bedrock AgentCore runtime instances are now generally available, letting teams run agents on their own EC2 instance types without managing the underlying infrastructure. Unlike the existing serverless microVM option, which caps sessions at 8 hours with faster startup, runtime instances support sustained sessions of up to 14 days and give access to GPU-accelerated, memory-optimized, and compute-optimized instance families for resource-intensive agents. The capability is available across nine regions in North America, Asia Pacific, and Europe, with AWS handling provisioning, patching, scaling, and lifecycle management while customers pay a management charge on top of EC2 costs. Source

2. Amazon Bedrock AgentCore adds temporal policies and rate limiting

AWS. Amazon Bedrock AgentCore now offers temporal policies and rate limiting to strengthen governance of production AI agents. Temporal policies evaluate stateful authorization rules against an agent’s session history rather than in isolation, enforcing workflow sequencing, requiring that a tool argument exactly match the output of a prior call, and mandating human approval for sensitive operations. Rate limiting applies per-user and per-group controls over requests to tools, models, and agents, including per-token inference limits and concurrent connection caps with OAuth and AWS IAM scoping, protecting downstream services and preventing multi-step attack chains. Source

3. Amazon ECS adds fractional GPU scheduling for small-model inference

AWS. Amazon ECS now supports fractional GPU scheduling on EC2 G6f instances, letting workloads run on GPU partitions as small as one-eighth of an NVIDIA L4 GPU with 3 GB of memory. Developers request fractional GPUs by setting GPU=0.125, 0.25, or 0.5 in task definitions, and ECS automatically places tasks on appropriately sized instances, with the feature explicitly positioned for small-model AI inference, model experimentation, and graphics rendering. By right-sizing containers instead of provisioning a full GPU, practitioners can cut inference costs while retaining automatic health monitoring and GPU metrics through CloudWatch across ECS Managed Instances and EC2 deployments. Source