BGL’s AI Agent: Revolutionizing Data Access with Claude & Bedrock

The article details BGL's implementation of an AI agent using Claude Agent SDK on Amazon Bedrock AgentCore to democratize business intelligence and overcome common data analysis bottlenecks. BGL, a financial services provider, previously struggled with non-technical users relying on data teams for insights and inconsistent results from traditional text-to-SQL solutions. Their new AI agent empowers employees to retrieve complex analytic insights through natural language queries, aligning with stringent financial services security and compliance requirements like session isolation and identity-based access controls.

A critical aspect of BGL's solution is a robust data foundation, built on Amazon Athena and dbt Labs. This system pre-processes terabytes of raw data into optimized, aggregated analytic tables, serving as a single source of truth. This “separation of concerns” ensures the AI agent focuses on interpreting natural language and generating simple SQL SELECT queries against these well-structured tables, significantly reducing hallucination and improving consistency, performance, and maintainability. For larger datasets and complex analysis, the agent writes and executes Python scripts directly on CSV results stored in its file system, effectively bypassing LLM context window limits.

The Claude Agent SDK provides key features like code execution for data processing and visualization, automatic context management for long-running sessions, and sandboxed execution for production-grade isolation. Its modular knowledge architecture uses `CLAUDE.md` for global project context and `SKILL.md` files for domain-specific expertise, organized by product lines. These skills are dynamically loaded, allowing the agent to progressively discover and apply relevant knowledge while maintaining efficient context usage.

Amazon Bedrock AgentCore hosts the Claude Agent SDK, offering fully-managed, stateful execution sessions lasting up to eight hours. Each session runs in an an isolated microVM, ensuring no cross-session data access and robust security. AgentCore's framework flexibility, alignment with security best practices, and integration capabilities were crucial for BGL. This architecture enables product managers, compliance teams, and customer success managers to access real-time, self-service analytics, transforming data access from a bottleneck into a competitive advantage and freeing data teams for strategic work.

(Source: https://aws.amazon.com/blogs/machine-learning/democratizing-business-intelligence-bgls-journey-with-claude-agent-sdk-and-amazon-bedrock-agentcore/)