Democratizing enterprise intelligence: BGL’s journey with Claude Agent SDK and Amazon Bedrock AgentCore

This submit is cowritten with James Luo from BGL.

Knowledge evaluation is rising as a high-impact use case for AI brokers. In response to Anthropic’s 2026 State of AI Agents Report, 60% of organizations rank information evaluation and report era as their most impactful agentic AI purposes. 65% of enterprises cite it as a high precedence. In observe, companies face two frequent challenges:

• Enterprise customers with out technical data depend on information groups for queries, which is time-consuming and creates a bottleneck.
• Conventional text-to-SQL options don’t present constant and correct outcomes.

Like many different companies, BGL confronted related challenges with its information evaluation and reporting use circumstances. BGL is a number one supplier of self-managed superannuation fund (SMSF) administration options that assist people handle the advanced compliance and reporting of their very own or a consumer’s retirement financial savings, serving over 12,700 companies throughout 15 international locations. BGL’s resolution processes advanced compliance and monetary information via over 400 analytics tables, every representing a particular enterprise area, comparable to aggregated buyer suggestions, funding efficiency, compliance monitoring, and monetary reporting. BGL’s prospects and staff want to search out insights from the information. For instance, Which merchandise had essentially the most damaging suggestions final quarter? or Present me funding tendencies for high-net-worth accounts. Working with Amazon Web Services (AWS), BGL constructed an AI agent utilizing Claude Agent SDK hosted on Amazon Bedrock AgentCore. Through the use of the AI agent enterprise customers can retrieve analytic insights via pure language whereas aligning with the safety and compliance necessities of economic companies, together with session isolation and identity-based entry controls.

On this weblog submit, we discover how BGL constructed its production-ready AI agent utilizing Claude Agent SDK and Amazon Bedrock AgentCore. We cowl three key points of BGL’s implementation:

• Why constructing a robust information basis is crucial for dependable AI agent-based text-to-SQL options
• How BGL designed its AI agent utilizing Claude Agent SDK for code execution, context administration, and domain-specific experience
• How BGL used AgentCore to supply the perfect stateful execution classes in manufacturing for a safer, scalable AI agent.

Establishing robust information foundations for an AI agent-based text-to-SQL resolution

When engineering groups implement an AI agent for analytics use circumstances, a typical anti-pattern is to have the agent deal with every thing together with understanding database schemas, reworking advanced datasets, checking out enterprise logic for analyses and deciphering outcomes. The AI agent is prone to produce inconsistent outcomes and fail by becoming a member of tables incorrectly, lacking edge circumstances, or producing incorrect aggregations.

BGL used its current mature huge information resolution powered by Amazon Athena and dbt Labs, to course of and rework terabytes of uncooked information throughout varied enterprise information sources. The extract, rework, and cargo (ETL) course of builds analytic tables and every desk solutions a particular class of enterprise questions. These tables are aggregated, denormalized datasets (with metrics and, summaries) that function a business-ready single supply of fact for enterprise intelligence (BI) instruments, AI brokers, and purposes. For particulars on how you can construct a serverless information transformation structure with Athena and dbt, see How BMW Group built a serverless terabyte-scale data transformation architecture with dbt and Amazon Athena.

The AI agent’s function is to deal with advanced information transformation inside the information system by specializing in deciphering the consumer’s pure language questions, translating it, and producing SQL SELECT queries towards well-structured analytic tables. When wanted, the AI agent writes Python scripts to additional course of outcomes and generate visualizations. This separation of considerations considerably reduces the chance of hallucination and affords a number of key advantages:

• Consistency: The information system handles advanced enterprise logic in a extra deterministic manner: joins, aggregations, and enterprise guidelines are validated by the information staff forward of time. The AI agent’s process turns into simple: interpret questions and generate fundamental SELECT queries towards these tables.
• Efficiency: Analytic tables are pre-aggregated and optimized with correct indexes. The agent performs fundamental queries fairly than advanced joins throughout uncooked tables, leading to a quicker response time even for giant datasets.
• Maintainability and governance: Enterprise logic resides within the information system, not within the AI’s context window. This helps be certain that the AI agent depends on the identical single supply of fact as different customers, comparable to BI instruments. If a enterprise rule modifications, the information staff updates the information transformation logic in dbt, and the AI agent routinely consumes the up to date analytic tables that replicate these modifications.

“Many individuals suppose the AI agent is so highly effective that they will skip constructing the information platform; they need the agent to do every thing. However you may’t obtain constant and correct outcomes that manner. Every layer ought to clear up complexity on the applicable stage” 

– James Luo, BGL Head of Knowledge and AI

How BGL builds AI brokers utilizing Claude Agent SDK with Amazon Bedrock

BGL’s improvement staff has been utilizing Claude Code powered by Amazon Bedrock as its AI coding assistant. This integration makes use of non permanent, session-based entry to mitigate credential publicity, and integrates with current identification suppliers to align with monetary companies compliance necessities. For particulars of integration, see Guidance for Claude Code with Amazon Bedrock

By way of its each day use of the Claude Code, BGL acknowledged that its core capabilities prolong past coding. BGL used its skill to motive via advanced issues, write and execute code, and work together with information and programs autonomously. Claude Agent SDK packages the identical agentic capabilities right into a Python and TypeScript SDK, in order that builders can construct customized AI brokers on high of Claude Code. For BGL, this meant they might construct an analytics AI agent with:

• Code execution: The agent writes and runs Python code to course of datasets returned from analytic tables and generate visualizations
• Computerized context administration: Lengthy-running classes don’t overwhelm token limits
• Sandboxed execution: Manufacturing-grade isolation and permission controls
• Modular reminiscence and data: A CLAUDE.md file for undertaking context and Agent Abilities for product line domain-specific experience

Why code execution issues for information analytics

Analytics queries usually return hundreds of rows and typically past megabytes of information. Commonplace tool-use, perform calling, and Mannequin Context Protocol (MCP) patterns usually cross retrieved information immediately into the context window, which shortly reaches mannequin context window limits. BGL applied a special strategy: the agent writes SQL to question Athena, then writes Python code to course of the CSV file outcomes immediately in its file system. This permits the agent to deal with massive end result units, carry out advanced aggregations, and generate charts with out reaching context window limits. You may be taught extra in regards to the code execution patterns in Code execution with MCP: Building more efficient agents.

Modular data structure

To deal with BGL’s various product traces and complicated area data, the implementation makes use of a modular strategy with two key configuration sorts that work collectively seamlessly.

CLAUDE.md (undertaking context)

The CLAUDE.md file offers the agent with world context—the undertaking construction, setting configuration (check, manufacturing, and so forth), and critically, how you can execute SQL queries. It defines which folders retailer intermediate outcomes and last outputs, ensuring information land in an outlined file path that customers can entry. The next diagram exhibits the construction of a CLAUDE.md file:

SKILL.md (Product area experience)

BGL organizes their agent area data by product traces utilizing the SKILL.md configuration information. Every talent acts as a specialised information analyst for a particular product. For instance, the BGL CAS 360 product has a talent referred to as CAS360 Knowledge Analyst agent, which handles firm and belief administration with ASIC compliance alignment; whereas BGL’s Easy Fund 360 product has a talent referred to as Easy Fund 360 Knowledge Analyst agent, which is supplied with SMSF administration and compliance-related area abilities. A SKILL.md file defines three issues:

• When to set off: What kinds of questions ought to activate this talent
• Which tables to make use of or map: References to the related analytic tables within the information folder (as proven within the previous determine)
• deal with advanced eventualities: Step-by-step steering for multi-table queries or particular enterprise questions if required

Through the use of SKILL.md information, the agent can dynamically uncover and cargo the correct talent to achieve domain-specific experience for corresponding duties.

• Unified context: When a talent is triggered, Claude Agent SDK dynamically merges its specialised directions with the worldwide CLAUDE.md file right into a single immediate. This permits the agent to concurrently apply project-wide requirements (for instance, all the time save to disk) whereas utilizing domain-specific data (comparable to mapping consumer inquiries to a bunch of tables).
• Progressive discovery: Not all abilities must be loaded into the context window without delay. The agent first reads the question to find out which talent must be triggered. It hundreds the talent physique and references to grasp which analytic desk’s metadata is required. It then additional explores corresponding information folders. This retains context utilization environment friendly whereas offering complete protection.
• Iterative refinement: If the AI agent is unable to deal with some enterprise data due to a scarcity of recent area data, the staff will collect suggestions from customers, determine the gaps, and add new data to current abilities utilizing a human-in-the-loop course of so abilities are up to date and refined iteratively.

As proven within the previous determine, agent abilities are organized per product line. Every product folder comprises a SKILL.md definition file and a references listing with extra area data and help supplies that the agent hundreds on demand.

For particulars about Anthropic Agent Abilities, see the Anthropic weblog submit, agents for the real world with Agent Skills

Excessive-level resolution structure

To ship a safer and scalable text-to-SQL expertise, BGL makes use of Amazon Bedrock AgentCore to host Claude Agent SDK whereas retaining information transformation within the current huge information resolution.

The previous determine illustrates a high-level structure and workflow. The analytic tables are pre-built each day utilizing Athena and dbt, and function the single supply of fact. A typical consumer interplay flows via the next phases:

1. Consumer request: A consumer asks a enterprise query utilizing Slack (for instance, Which merchandise had essentially the most damaging suggestions final quarter?).
2. Schema discovery and SQL era: The agent identifies related tables utilizing abilities and writes SQL queries.
3. SQL safety validation: To assist forestall unintended information modification, a safety layer permits solely SELECT queries and blocks DELETE, UPDATE, and DROP operations.
4. Question execution: Athena executes the question and shops outcomes into Amazon Simple Storage Service (Amazon S3).
5. End result Obtain: The agent downloads the ensuing CSV file to the file system on AgentCore, utterly bypassing the context window to keep away from token limits.
6. Evaluation and visualization: The agent writes Python code to investigate the CSV file and generate visualizations or refined datasets relying on the enterprise query.
7. Response supply: Last insights and visualizations are formatted and returned to the consumer in Slack.

Why use Amazon Bedrock AgentCore to host Claude Agent SDK

Deploying an AI agent that executes arbitrary Python code requires vital infrastructure concerns. As an example, you want isolation to assist be certain that there’s no cross-session entry to information or credentials. Amazon Bedrock AgentCore offers fully-managed, stateful execution classes, every session has its personal remoted microVM with a separate CPU, reminiscence, and file system. When a session ends, the microVM terminates absolutely and sanitizes reminiscence, serving to to make sure no remnants persist for future classes. BGL discovered this service particularly useful:

• Stateful execution session: AgentCore maintains session state for as much as 8 hours. Customers can have ongoing conversations with the agent, referring again to earlier queries with out dropping context.
• Framework flexibility: It’s framework-agnostic. It helps deployment of AI brokers comparable to Strands Agents SDK, Claude Agent SDK, LangGraph, and CrewAI with just a few traces of code.
• Aligned with safety finest practices: It offers session isolation, VPC help, AWS Identity and Access Management (IAM) or OAuth primarily based identification to facilitate ruled, compliance-aligned agent operations at scale.
• System integration: This can be a forward-looking consideration.

“There’s Gateway, Reminiscence, Browser instruments, an entire ecosystem constructed round it. I do know AWS is investing on this path, so every thing we construct now can combine with these companies sooner or later.”

– James Luo, BGL Head of Knowledge and AI. 

BGL is already planning to combine AgentCore Reminiscence for storing consumer preferences and question patterns.

Outcomes and impression

For BGL’s greater than 200 staff, this represents a big shift in how they extract enterprise intelligence. Product managers can now validate hypotheses immediately with out ready for the information staff. Compliance groups can spot danger tendencies with out studying SQL. Buyer success managers can pull account-specific analytics in real-time throughout consumer calls. This democratization of information entry helps rework analytics from a bottleneck right into a aggressive benefit, enabling quicker decision-making throughout the group whereas releasing the information staff to give attention to strategic initiatives fairly than one-time question requests.

Conclusion and key takeaways

BGL’s journey demonstrates how combining a powerful information basis with agentic AI can democratize enterprise intelligence. Through the use of Amazon Bedrock AgentCore and the Claude Agent SDK, BGL constructed a safer and scalable AI agent that empowers staff to faucet into their information to reply enterprise questions. Listed here are some key takeaways:

• Put money into a powerful information basis: Accuracy begins with a powerful information basis. Through the use of the information system and information pipeline to deal with advanced enterprise logic (joins and aggregations), the agent can give attention to fundamental, dependable logic.
• Arrange data by area: Use Agent Abilities to encapsulate domain-specific experience (for instance, Tax Legislation or Funding Efficiency). This retains the context window clear and manageable. Moreover, set up a suggestions loop: constantly monitor consumer queries to determine gaps and iteratively replace these abilities.
• Use code execution for information processing: Keep away from utilizing an agent to course of massive datasets utilizing a big language mannequin (LLM) context. As an alternative, instruct the agent to write down and execute code to filter, combination, and visualize information.
• Select stateful, session-based infrastructure to host the agent: Conversational analytics requires persistent context. Amazon Bedrock AgentCore simplifies this by offering built-in state persistence (as much as 8-hour classes), assuaging the necessity to construct customized state dealing with layers on high of stateless compute.

For those who’re able to construct related capabilities on your group, get began by exploring the Claude Agent SDK and a brief demo of Deploying Claude Agent SDK on Amazon Bedrock AgentCore Runtime. You probably have an analogous use case or want help designing your structure, attain out to your AWS account staff.

References:

Concerning the authors

Dustin Liu is a options architect at AWS, targeted on supporting monetary companies and insurance coverage (FSI) startups and SaaS firms. He has a various background spanning information engineering, information science, and machine studying, and he’s obsessed with leveraging AI/ML to drive innovation and enterprise transformation.

Melanie Li, PhD, is a Senior Generative AI Specialist Options Architect at AWS primarily based in Sydney, Australia, the place her focus is on working with prospects to construct options leveraging state-of-the-art AI and machine studying instruments. She has been actively concerned in a number of Generative AI initiatives throughout APJ, harnessing the facility of Massive Language Fashions (LLMs). Previous to becoming a member of AWS, Dr. Li held information science roles within the monetary and retail industries.

Frank Tan is a Senior Options Architect at AWS with a particular curiosity in Utilized AI. Coming from a product improvement background, he’s pushed to bridge know-how and enterprise success.

James Luo is Head of Knowledge & AI at BGL Company Options, a world-leading supplier of compliance software program for accountants and monetary professionals. Since becoming a member of BGL in 2008, James has progressed from developer to architect to his present management function, spearheading the Knowledge Platform and Roni AI Agent initiatives. In 2015, he shaped BGL’s BigData staff, implementing the primary deep studying mannequin within the SMSF business (2017), which now processes 200+ million transactions yearly. He has spoken at Massive Knowledge & AI World and AWS Summit, and BGL’s AI work has been featured in a number of AWS case research.

Dr. James Bland is a Know-how Chief with 30+ years driving AI transformation at scale. He holds a PhD in Laptop Science with a machine studying focus and leads strategic AI initiatives at AWS, enabling enterprises to undertake AI-powered improvement lifecycles and agentic capabilities. Dr. Bland spearheaded the AI-SDLC initiative, authored complete guides on Generative AI within the SDLC, and helps enterprises architect production-scale AI options that basically rework how organizations function in an AI-first world.