AWS AI Updates: June 4, 2026

1. SageMaker AI Launches Multi-Turn Reinforcement Learning for Agent Customization

AWS. Amazon SageMaker AI added a serverless multi-turn reinforcement learning capability that fine-tunes foundation models against a customer’s own agent environment, rewarding the full sequence of decisions an agent makes across a task rather than single responses. The service manages rollout orchestration, trajectory collection, training, and checkpointing, with built-in MLflow tracking and evaluation metrics including reward, pass@k, and trajectory measurements. Supported models include Qwen 3.6 27B, Nova Lite 2.0, GPT-OSS-20B, and Gemma 31B, with agents able to run on Amazon Bedrock AgentCore Runtime, EKS, EC2, or Fargate, and billing based only on tokens processed. Source

2. AWS Step Functions Adds an AgentCore-Powered Agentic Reasoning Step

AWS. AWS Step Functions now offers an optimized integration with the managed harness in Amazon Bedrock AgentCore, letting developers embed AI agent reasoning steps directly into orchestrated workflows. Practitioners can automate tasks such as document classification and form data extraction, run multiple agents in parallel or sequentially, insert human approval gates, and customize the model, system prompt, and tools per invocation without duplicating configurations. The step supports session IDs for persisting agent context across invocations and surfaces full execution visibility including agent input/output, token usage, and CloudWatch links, available in US East (N. Virginia), US West (Oregon), Europe (Frankfurt), and Asia Pacific (Sydney). Source

3. Amazon ECS Managed Instances Now Supports AWS Trainium and Inferentia

AWS. Amazon ECS Managed Instances now supports AWS Trainium and AWS Inferentia accelerators, including Inferentia2, Trainium1, and Trainium2 instance types, giving teams fully managed compute for generative AI training and inference without provisioning underlying infrastructure. Users create a capacity provider with the desired accelerator instance types and add NEURON_CORE=all to the task definition’s ResourceRequirement section, which directs ECS to launch the specified instances and place a single task per instance with all accelerator resources allocated to that workload. The feature can be enabled through the AWS Console, the Amazon ECS MCP Server, or infrastructure-as-code tools on new or existing clusters. Source

4. SageMaker Data Agent Adds Conversation History

AWS. Amazon SageMaker Data Agent in SageMaker Unified Studio now maintains conversation history, letting analysts access a scrollable list of past conversations via a clock icon in the chat panel and resume multi-step analyses without losing context. The feature auto-generates conversation titles and timestamps, preserving previously generated code and troubleshooting steps across notebooks and Query Editor workflows so teams can move between concurrent projects without rebuilding prior work. It is available in all AWS regions where SageMaker Data Agent currently operates. Source

5. SageMaker Studio Quick Setup Ships With Model Customization Ready by Default

AWS. Amazon SageMaker Studio reduced its quick setup time from over two minutes to under twenty seconds while preconfiguring serverless model customization permissions out of the box. A managed policy named AmazonSageMakerModelCustomizationCoreAccess is created and attached by default, granting permissions for fine-tuning with custom reward functions for reinforcement learning, model evaluation, and deployment, removing manual IAM configuration before a first run. The capability is available across all AWS commercial regions that support SageMaker Studio, with existing environments receiving actionable messages and documentation links to enable it. Source