Stripe Kai: Deep Agents, 1,000 Skills and 83% Weekly Use
Agentic AIOne Stripe engineer built Kai, the company’s internal AI agent, in a week, and 83% of Stripe now uses it every week. The harder lesson sits further down LangChain’s case study: frontier models got worse once Kai held more than 150 skills alongside its system prompt. For platform teams building an internal agent for non-engineers, Stripe’s architecture is a working answer to both the speed question and the scale problem.
How It Works
LangChain’s case study describes Kai, Stripe’s Knowledge AI Platform, as “a coding agent for non-engineers.” Users chat in a session, and Kai produces reports, dashboards or documents that evolve alongside the conversation, drawing on Stripe’s data warehouse, Slack and Google Suite.
The stack has four layers. Deep Agents, LangChain’s open-source agent harness, handles the tool-calling loop, middleware, streaming and state. A Stripe-specific harness above it ties Kai into Stripe’s security and internal services. A configuration layer lets teams define their own Kai agents with different skills and personas, and the Kai UI sits on top. Engineering manager Sharadh Krishnamurthy put the split this way: “The Deep Agents layer solves all the non-Stripey problems, so that we can focus on solving the Stripey agent problems.”
Three pieces of middleware made the one-week build possible. A virtual filesystem backed by S3 persists files across turns, with a “sync in / sync out” pattern around every sandbox call. The sandbox is exposed to the agent as a tool, not as the agent’s own runtime, so the agent runs outside it. Summarization middleware keeps long, multi-turn sessions inside context limits, with tuning for threshold, summarizer model and output size.
Skills do the routing. Stripe has more than 500 internal MCP tools and over 1,000 skills from more than 100 teams. Each skill’s allowedTools list drives dynamic loading, so choosing a skill decides which tools enter context, rather than loading every tool upfront. Foundational skills are pinned so Stripe context and policy always stay loaded, and extra skills load by user profile, so a sales ops user gets a different set than finance.
What This Means for You
Adoption came from the people the team designed for. In open preview, Kai hit its quarterly adoption target in a week and grew from 296 users to over 5,000 in roughly four weeks. Today it runs more than 60,000 sessions, with higher usage in Marketing (95%) and go-to-market teams (87%) than in engineering.
The limits are just as instructive. With a 1,024-character limit on skill frontmatter, the team saw quality fall past 150 skills, and LangChain says the skill count “remains a challenge.” Stripe’s plan is hybrid selection: pure LLM selection works at current scale, but a RAG or classifier layer will prefilter before the LLM decides. Stack Archive’s look at Deep Agents’ context offloading and summarization covers the same middleware from the context side.
The caveat is the source: this is a vendor case study, and the adoption figures are Stripe’s own, reported by LangChain.
The So What
Teams building an internal agent should copy two of Stripe’s choices early: let skill selection gate which tools load, and keep code execution in a sandbox the agent calls as a tool. Then measure answer quality as the skills library grows, and plan a prefilter before it reaches the roughly 150-skill point where Stripe saw frontier models degrade.
Source: LangChain Blog
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →