How Stripe Built Kai on Deep Agents in 1 Week

Deep Agents abstracts the ‘non-domain’ agent infrastructure, allowing engineers to focus on specific business workflows.

Key Points

Stripe developed ‘Kai’, a company-wide productivity agent, utilizing the LangChain and LangGraph stack, specifically leveraging the ‘Deep Agents’ open-source harness.
The architecture is layered: Deep Agents (base primitives) $\to$ Stripe-specific harness (security/infra) $\to$ Configuration layer (custom personas/skills) $\to$ Kai UI.
Production-readiness was achieved through three key middleware components: a virtual filesystem (S3-backed) for persistent session context, a sandboxed environment for secure code execution (analytics/PDF processing), and summarization middleware to manage long-turn context limits.
Kai employs a federated ‘skills’ model with over 1,000 skills from 100+ teams, using a two-pass dynamic tool loading system to prevent model quality degradation caused by oversized system prompts.
The project validated Stripe’s investment in a Python-native stack, as a single engineer built the initial version in one week using Deep Agents’ primitives.
Adoption grew from ~300 to 5,000+ users in four weeks, with particularly high penetration in non-engineering functions like Marketing (95%) and GTM (87%).