How Stripe Built Kai, its Company-Wide AI Agent, on Deep Agents
- Source
- langchain.com
- Author
- LangChain
- Date
Context
Stripe's employee assistant Kai selects task instructions before loading the tools those instructions require. In LangChain's case study, teams maintain skills that describe how to perform particular work and identify the relevant tools. Kai starts with skills appropriate to the employee and configuration, then uses selection to bring additional tools into the model's working .
This makes a large internal tool collection manageable without presenting every tool on every request. Some foundational skills stay loaded. The team reports degraded quality when its particular prompt included more than 150 skill descriptions, motivating further work on selection. That observation concerns Kai's configuration, rather than establishing a universal limit for language models.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.
- context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
- agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
“With Claude Code’s launch in late 2024, suddenly every Stripe employee wanted to build agents. But for non-engineers, the terminal, data access, and security components were significant barriers.”
LangChain
“The Deep Agents layer solves all the non-Stripey problems, so that we can focus on solving the Stripey agent problems," explained Sharadh. "You can take a bunch of middleware off the shelf. The composable skill patterns, the composite backend, it provides the right structure, but still enough flexibility.”
LangChain
“With 500+ internal MCP tools available and a growing skills library, Kai cannot load everything into context. Agent Skills ’ allowedTools lists drive dynamic tool loading, creating a two-pass system where skill selection gates tool context, rather than loading all tools upfront.”
LangChain
“With a 1024 character limit on frontmatter, the team discovered quality degradation of frontier models with more than 150 skills when combined with their system prompt.”
LangChain
“Today, 83% of Stripe uses Kai every week, across >60k sessions. Kai sees more adoption by percentage from business functions like Marketing (95%) and GTM teams (87%) than even engineering”
LangChain
Checking sign-in…
Loading comments…