aws AWS Big Data Blog ·

AWS introduces architectural patterns for agentic AI with real-time data

blogaiawsarchitectmedia
announcement

AWS shares three architectural patterns to power agentic AI applications using real-time streaming data. These patterns address the evolving landscape of generative AI, where systems observe, reason, and act autonomously in production. The guidance helps developers build continuous data flows for real-time inference and action, event-driven agent invocation, and context synchronization. This approach utilizes existing AWS services like Amazon MSK, Kinesis, Flink, Bedrock, and SageMaker to keep agent memory current and lakehouses fresh for training.

  • Understanding the Shift to Agentic AI Applications
  • Streaming Feature Engineering for Real-time Inference and Action
  • Event-Driven Agent Invocation for Proactive Responses
Notes (3)
  • Understanding the Shift to Agentic AI Applications

    Generative AI systems now observe, reason, and act autonomously, requiring architectures where streaming data continuously powers agent action and keeps real-time lakehouses fresh for training and retrieval. This moves beyond simply injecting context into large language models, demanding new relationships between streaming infrastructure and AI.

  • Streaming Feature Engineering for Real-time Inference and Action

    This pattern combines real-time data for continuous AI model improvement and immediate action. Streaming events are processed by Amazon Managed Service for Apache Flink to produce features for both inference via Amazon Bedrock or SageMaker, and continuous training by writing to S3 Tables for fresh datasets, enabling models to be retrained on data minutes old.

  • Event-Driven Agent Invocation for Proactive Responses

    This pattern involves streaming pipelines continuously analyzing data and invoking agentic workflows when specific conditions or patterns are met, acting as a sensor for agents. Amazon Managed Service for Apache Flink detects anomalies, assembles a full context package including historical and correlated data, and publishes it for consumption by an Amazon Bedrock AgentCore agent.

Read the original announcement →

https://aws.amazon.com/blogs/big-data/powering-agentic-ai-with-real-time-streaming-data-on-aws/

Related releases