Building AI agents for financial services involves navigating a complex landscape that demands precision. The author shares insights from developing Fintool, an AI agent designed to meet the rigorous standards of professional investors. Key lessons include the necessity of a sandbox environment for multi-step workflows, as executing arbitrary code poses significant risks. For instance, a misstep from an LLM led to a catastrophic command that could have wiped data, underscoring the importance of isolating user operations. Each user now operates in their own sandbox, leveraging AWS Attribute-Based Access Control to ensure data privacy and security. Contextual data management is equally vital. The diverse nature of financial data—from SEC filings to earnings transcripts—requires normalization into structured formats for effective use by the AI. The author emphasizes the transformation of messy, heterogeneous data into clean, searchable context, which involves converting everything into markdown for narratives, CSV for structured data, and JSON for metadata. The chunking strategy is crucial; it determines how data is retrieved and processed by the agent. For example, 10-K filings are segmented by regulatory structure, while news articles are chunked at the paragraph level. This meticulous approach to data management ensures the agent provides accurate and relevant responses to user queries. The author also highlights the importance of rigorous evaluation and monitoring. Professional investors rely on the AI for high-stakes decisions, meaning even a small error can lead to a loss of trust. The AI's outputs must be double-checked and validated, with domain-specific evaluations in place to catch potential mistakes before they impact financial decisions. The lessons learned from building Fintool reveal a commitment to detail and a willingness to adapt, which are essential for success in the demanding field of financial services.