Overview

Ramp’s engineering team shares lessons from building AI agents for finance operations, covering their journey from building hundreds of separate agents to consolidating into a single agent with thousands of skills. They detail how they built their policy agent, the infrastructure required, and the cultural shifts needed for AI-native product development.

Key Takeaways

  • Start simple and iterate quickly - Begin with constrained problems like coffee expenses rather than trying to automate all of finance from day one
  • Build your own ground truth dataset through cross-functional labeling sessions - Users are often wrong about policy decisions, so you need your own definition of correctness
  • Consolidate to single agent architecture - Instead of building thousands of separate agents, focus on one agent with thousands of skills and shared toolboxes
  • Design for auditability from the beginning - As systems become more complex and black-box-like, assume you only know inputs/outputs and ensure you can verify correctness
  • Cultural shift required: focus on impact over coding - Teams that understand users, handle ambiguity, and obsess over experience will outperform those who debate libraries and bike-shed details

Topics Covered