Stripe builds Kai AI agent on Deep Agents in one week

Stripe's AI platform team faced a common problem as agentic AI adoption accelerated: engineering teams across the company were building orchestration layers on their own Ruby and Java stacks, but production-quality agent harnesses required performance, evaluation and reliability work beyond simple connectivity. When Claude Code launched in late 2024, every Stripe employee wanted to build agents, but non-engineers hit barriers around the terminal, data access and security. Sharadh Krishnamurthy and Anupam Upadhyay inverted the approach, meeting non-technical employees where they were with a purpose-built assistant rather than pushing them toward developer tooling.

The resulting Knowledge AI Platform, Kai, sits on a layered architecture: Deep Agents as the base handling LLM interaction primitives, a Stripe-specific harness integrating security and internal services, a configuration layer for custom agents, and the Kai UI. Sharadh described the Deep Agents layer as solving all the non-Stripey problems so the team could focus on Stripey agent problems. Anupam built the initial version in one week, and that speed demonstrated the return on Stripe's investment in rebuilding internal service scaffolding around a Python-native stack after more than a decade on Ruby and Java.

Kai's early traction was rapid: it hit the quarterly adoption target in a week and grew more than 16x from 296 users to over 5,000 in roughly four weeks. The bulk of new users were sellers, finance analysts and business operations people who had been told to use AI but had not found a tool that fit how they worked. Today 83% of Stripe uses Kai every week across more than 60k sessions, with Marketing at 95% and GTM teams at 87% adoption by percentage. The team now faces scaling challenges around skill selection, governance and personalization as it moves past product-market fit.