Stripe built its company agent in one week and 83% of the company now uses it weekly
One engineer. One week. Then 5,000 users in four weeks, a 16x jump that cleared the quarterly target in the first seven days. Kai is now used weekly by 83% of Stripe across more than 60,000 sessions.
Kai is Stripe's Knowledge AI Platform β a company-wide agent for the work that isn't coding. You start a session, you chat, and it produces artifacts: reports, dashboards, documents that live next to the conversation and keep evolving as you talk. Stripe's own framing is the cleanest description of the category: it's a coding agent for people who don't write code.
It's built on Deep Agents, LangChain's open-source harness, and the architecture is four layers stacked in a way worth stealing. Deep Agents at the base handling LLM primitives, request management, middleware. A Stripe-specific harness layer wiring in security, infrastructure, and internal services. A configuration layer that spins up agent instances with particular skill sets. Then the Kai UI on top. The middleware is where the production work lives: an S3-backed virtual filesystem so file context survives across turns, a sandbox for Python analytics and document processing, and summarization middleware to keep long sessions from eating the budget.
The scale numbers are the ones people building this internally should write down. 500+ internal MCP tools. 1,000+ skills contributed by more than 100 teams. And the failure mode they hit: quality degrades past roughly 150 skills loaded alongside the system prompt. Their fix is two-pass skill loading, where the model picks relevant skills instead of everything being present. That ceiling is the most useful thing in the whole writeup, because everyone building a company agent is heading straight for it and most will discover it the hard way.
Adoption splits are worth noting too: Marketing 95% weekly, GTM 87%, and engineers using it as a copilot despite not being the target. One employee's line was "my Stripe career is divided into before and after Kai," which is the kind of quote a comms team writes, except the 60,000 sessions make it hard to dismiss.
https://stripe.dev/blog/meet-stripes-knowledge-ai-platform
← Back to all articles
Kai is Stripe's Knowledge AI Platform β a company-wide agent for the work that isn't coding. You start a session, you chat, and it produces artifacts: reports, dashboards, documents that live next to the conversation and keep evolving as you talk. Stripe's own framing is the cleanest description of the category: it's a coding agent for people who don't write code.
It's built on Deep Agents, LangChain's open-source harness, and the architecture is four layers stacked in a way worth stealing. Deep Agents at the base handling LLM primitives, request management, middleware. A Stripe-specific harness layer wiring in security, infrastructure, and internal services. A configuration layer that spins up agent instances with particular skill sets. Then the Kai UI on top. The middleware is where the production work lives: an S3-backed virtual filesystem so file context survives across turns, a sandbox for Python analytics and document processing, and summarization middleware to keep long sessions from eating the budget.
The scale numbers are the ones people building this internally should write down. 500+ internal MCP tools. 1,000+ skills contributed by more than 100 teams. And the failure mode they hit: quality degrades past roughly 150 skills loaded alongside the system prompt. Their fix is two-pass skill loading, where the model picks relevant skills instead of everything being present. That ceiling is the most useful thing in the whole writeup, because everyone building a company agent is heading straight for it and most will discover it the hard way.
Adoption splits are worth noting too: Marketing 95% weekly, GTM 87%, and engineers using it as a copilot despite not being the target. One employee's line was "my Stripe career is divided into before and after Kai," which is the kind of quote a comms team writes, except the 60,000 sessions make it hard to dismiss.
https://stripe.dev/blog/meet-stripes-knowledge-ai-platform
Comments