AWS re:Invent 2025 - Autonomous agents powered by streaming data and Retrieval Augmented Generation

Autonomous Agents Powered by Streaming Data and Retrieval Augmented Generation

Key Challenges Faced by IT and Network Operations

• Sudden spikes in network traffic (e.g. 600% increase) with high-severity anomalies detected
• Flood of log entries and alerts making it difficult to identify the root cause
• Lack of context and historical information to understand and resolve issues
• Alert overload and "alert fatigue" leading to potential oversight of critical problems
• Siloed data across multiple systems (logs, metrics, traces) making correlation difficult
• Manual troubleshooting and investigation taking hours or days to resolve incidents

Leveraging Streaming Data, Intelligent Agents, and Knowledge Bases

• Streaming services (Amazon MSK, Amazon Kinesis) to capture and store high-velocity data reliably
• Apache Flink for continuous stream processing to enrich and transform the data
• Amazon OpenSearch Service as a scalable, low-latency vector database for agent knowledge bases
• AWS Bedrock for managed agent creation, leveraging large language models and retrieval-augmented generation
• Strand Agents SDK for developers to build custom agents integrated with AWS Agent Core

Agentic Reasoning and Action Loop

1. Agent receives a problem statement or prompt
2. Agent uses large language models to understand the context and reason about the issue
3. Agent identifies relevant tools and data sources to gather more information
4. Agent orchestrates the execution of these tools using the MCP (Model Context Protocol)
5. Agent iterates through this loop, gathering more context and refining its understanding
6. Agent provides a final result or recommended actions to address the problem

Bedrock Agents and Knowledge Bases

• Bedrock Agents provide a fully managed service for creating and hosting intelligent agents
• Agents can leverage Bedrock's knowledge bases, built using models like Titan Embedding V2
• Knowledge bases store contextual information from runbooks, security standards, and other sources
• Agents use this knowledge to enhance their reasoning and provide more informed recommendations

Strand Agents and Agent Core

• Strand Agents SDK allows developers to build custom agents using Python or TypeScript
• Agent Core provides a serverless runtime, identity management, and tool integration capabilities
• Agents can leverage the MCP protocol to communicate with various tools and data sources

Business Impact and Use Cases

• IT and Network Operations: Proactive monitoring, faster incident resolution, reduced downtime
• Manufacturing: Predictive maintenance, safer operations, reduced unplanned downtime
• Automotive: Automated service scheduling, emergency assistance, personalized routing
• Healthcare: Remote patient monitoring, automated appointment scheduling, emergency response

Key Takeaways

• Real-time data is powerful when combined with contextual understanding from the past
• Intelligent agents can bridge the gap between streaming data and organizational knowledge
• Agents can automate reasoning, decision-making, and execution of actions on behalf of humans
• Flexible agent frameworks (Bedrock, Strand) allow for both managed and custom-built solutions
• Agents can be applied across industries to enable proactive, autonomous, and personalized systems

Resources

• AWS Bedrock: https://aws.amazon.com/bedrock/
• AWS Agent Core: https://aws.amazon.com/agent-core/
• Strand Agents SDK: https://aws.amazon.com/solutions/implementations/strand-agents/
• Anomaly Detection Workshop: https://github.com/aws-samples/aws-reinvent-2025-anomaly-detection-workshop