How Stripe Built Kai with Deep Agents in a Week

Stripe built Knowledge AI Platform (Kai), a company-wide productivity agent that any employee can use. Kai connects to internal data warehouses, Slack, and Google Suite to synthesize data, write documents, analyze trends, and generate dashboards and reports.

Unlike typical AI assistants, Kai is designed to understand Stripe's internal systems and work norms. It leverages over 1,000 skills contributed by more than 100 teams, so users don't need to explain their role or company context every time.

To let non-engineer employees use the agent without the barriers of developer tools like terminals, data access, or security configuration, Stripe provided a session-based chat interface along with a workspace where outputs evolve alongside the conversation. Users can continue editing artifacts, reports, dashboards, and documents generated during the conversation.

Kai's foundation is Deep Agents, LangChain's open-source agent harness. Because Deep Agents handles common functionality such as LLM request management, agent execution, tool call loops, middleware composition, streaming, and state management, the Stripe team was able to focus on their own workflows and security instead of building infrastructure from scratch.

Stripe's layered structure is as follows.

• Deep Agents: the foundational layer for LLM interaction and agent execution
• Stripe-specific harness: the middle layer that integrates security policies, infrastructure, and internal services
• Configuration layer: the layer where developers configure Kai agents with different skills, behaviors, and personas without modifying the base harness
• Kai UI: the product interface that employees directly use

Stripe's initial version of Kai was built in just one week. The team leveraged Deep Agents' middleware and composable skills and backend structure to quickly secure the foundational functionality needed for a production environment.