SD Biosensor Productionizes Conversational Data Analysis Agent with Amazon Bedrock AgentCore

SD Biosensor transitioned its conversational data analysis agent to a production environment using Amazon Bedrock AgentCore and the Strands Agents SDK through the AWS 3A program. The architecture was redesigned to address issues from the initial prototype, including lack of reproducibility, increased latency and costs, and limited observability.

Transition from Multi-Agent to Single-Agent

The initial structure was a multi-agent system comprising an orchestrator and multiple sub-agents, but it suffered from latency and token costs due to context round-trips. To resolve this, the architecture was simplified to a single-agent structure, and calculations requiring determinism were moved to version-pinned pure function tools. The LLM is responsible only for tool selection and result summarization, while code performs value calculations to ensure reproducibility.

Separation of Code Execution Lanes

Code execution environments were separated into three lanes based on task nature. Critical figures requiring validation are calculated immediately via deterministic tools within the agent process. Lightweight tasks such as free-form graphs are executed in the managed sandbox of Amazon Bedrock AgentCore Code Interpreter. Long-running tasks like large-scale data processing or ML training are handled asynchronously via Amazon SageMaker Processing jobs. Routing is determined by predefined scope tables rather than model judgment.

Observability and Asynchronous Processing

Execution flows are instrumented with the OpenTelemetry standard and sent to Amazon Bedrock AgentCore Observability and Langfuse. Traces are forcibly flushed at the end of each turn to prevent data loss. Completion notifications for asynchronous tasks such as SageMaker Processing are loaded into a DynamoDB inbox via Amazon EventBridge and Lambda, allowing users to view results even when not connected. Scheduled tasks are managed via EventBridge Scheduler.