Stripe Built Its Company-Wide AI Agent Kai in One Week

Kai brings Stripe context to employee workflows

Stripe’s Knowledge AI Platform, known as Kai, is available to every employee. The session-based interface connects to Stripe’s internal data warehouse, Slack and Google Suite, allowing users to synthesize data, brainstorm, draft documents, analyze trends and collaborate across functions.

Kai can produce artifacts, reports, dashboards and documents alongside the conversation. Those outputs can evolve as a session continues, giving the system a workflow similar to a coding agent, but designed for non-engineers.

Unlike a generic assistant, Kai is preloaded with knowledge of Stripe’s internal systems, data sources and norms. More than 100 teams contribute to a library containing over 1000 skills. These skills provide domain-specific instructions, tools and approaches for different kinds of work.

Deep Agents supplied the underlying infrastructure

Stripe’s AI teams had previously built orchestration layers on top of existing Ruby and Java stacks. As more employees wanted to create agents, the company also faced the challenge of making agent development accessible beyond engineering teams.

Stripe built Kai in layers. Deep Agents provides the base for LLM interaction, agent execution, request management and middleware composition. A Stripe-specific harness adds the company’s security posture, infrastructure and internal services. Teams can then configure specialized Kai agents with different skills, behaviors and personas, while the Kai interface connects employees to the full stack.

“The Deep Agents layer solves all the non-Stripey problems, so that we can focus on solving the Stripey agent problems,” Sharadh Krishnamurthy, Engineering Manager for Stripe’s Agent Foundation team, said in the LangChain case study.

Middleware helped make the system production-ready

Stripe used filesystem middleware to create a virtual filesystem backed by S3. Files can be read, written and referenced across turns, while a “sync in / sync out” process keeps files available before and after sandbox execution.

Sandbox middleware supports analytics, including Python-based data queries and charts, as well as processing for PDFs, presentations and structured documents. The agent calls the sandbox as a tool, while the agent itself runs outside that environment.

Summarization middleware helps Kai manage long, multi-turn sessions. Stripe adjusts settings such as the summarization threshold, summarizer model and output size to balance context, performance and cost.

The result also demonstrated the value of Stripe’s investment in a Python-native stack. According to LangChain, one engineer built Kai’s initial version in one week, while Deep Agents supplied much of the agent infrastructure that would otherwise have required months to build and harden.