Editorial illustration for Stripe Built Kai on Deep Agents in 1 Week to Navigate Hundreds of Internal Tools
Stripe Built AI Agent Kai in One Week Using Deep Agents
Stripe Built Kai on Deep Agents in 1 Week to Navigate Hundreds of Internal Tools
Stripe's AI platform team built Kai, the company's internal "Knowledge AI" agent, in a single week using LangChain's Deep Agents framework. The tool now lets employees navigate hundreds of internal systems through plain conversation instead of hunting through Stripe's Ruby and Java stacks by hand.
The project grew out of a problem that surfaced across Stripe's engineering org once Claude Code shipped in late 2024. Suddenly, according to the team, every employee wanted an agent of their own, but most of them weren't engineers. They couldn't work a terminal, didn't have the access permissions, and had no way to touch the security layers that sat between them and Stripe's internal data.
Sharadh Krishnamurthy, Engineering Manager for Stripe's Agent Foundation team, and Staff Software Engineer Anupam Upadhyay were watching teams across the company build their own orchestration layers on top of existing infrastructure. Connecting to internal systems was manageable. Making those agents reliable at scale was not. Krishnamurthy and Upadhyay landed on a different starting point: build for the people who'd never open a terminal in the first place.
One engineer built Kai in a single week. The primitives from Deep Agents meant that the hard agent infrastructure was already solved. That speed proved that the time spent building internal Stripe support for Python was paid back immediately.
Why this matters
A one-week build is the headline, but the real signal is the skills architecture underneath it. Stripe didn't hand Kai a giant prompt and hope it could juggle hundreds of internal tools. It built modular, agent-executable units that each encapsulate a task, its required tools, and the approach to solving it. That's a structural choice, not a scale trick, and it's the part worth copying.
For teams building internal agents, this is a useful template: the bottleneck isn't getting a model to call tools, it's organizing institutional knowledge into pieces an agent can load on demand. Stripe's AI platform team treating this as shared infrastructure, rather than a one-off bot, suggests they're thinking about maintainability past the demo stage.
We'd want to know how skills get authored and updated as Stripe's internal tools change, and whether this pattern holds at companies with messier, less documented systems than a payments giant with disciplined engineering culture. The one-week build number is catchy. The skills abstraction is what determines if Kai still works in a year.
Related Reading
AI Tools & Apps Claude gains shared context in Excel, PowerPoint; Microsoft adds Copilot Cowork
AI Tools & Apps Sarvam Signs 5-Year Defence AI Deal Covering Army, Navy and Air Force
AI Tools & Apps LG's recent webOS update adds Microsoft Copilot app, now removable
From Archives Indie gaming reckons after House House admits using AI to fix Big Walk bug
From Archives Copilot Caught 8 Diff Bugs With Rules, Caught 0 of 2 Outside the Diff
Common Questions Answered
How did Stripe build Kai using LangChain's Deep Agents framework in just one week?
Stripe's AI platform team leveraged LangChain's Deep Agents framework, which provided pre-built agent infrastructure primitives that eliminated the need to build foundational agent systems from scratch. By using these ready-made components, a single engineer was able to construct Kai's core functionality in just seven days, demonstrating how the framework's modular design accelerated development significantly.
What problem did Kai solve for Stripe employees navigating internal tools?
Before Kai, Stripe employees had to manually hunt through hundreds of internal systems built on Ruby and Java stacks to find information. Kai transformed this experience by allowing employees to navigate these internal tools through plain conversation instead, eliminating the need for direct knowledge of the underlying technical infrastructure.
Why did demand for an internal agent like Kai increase after Claude Code shipped in late 2024?
The release of Claude Code in late 2024 demonstrated the capabilities of AI agents to Stripe's engineering organization, which sparked widespread interest among employees who suddenly wanted their own agent. This surge in demand prompted Stripe's AI platform team to quickly build Kai as a scalable solution to meet this organizational need.
What structural approach did Stripe use to make Kai work with hundreds of internal tools?
Rather than giving Kai a single large prompt and hoping it could manage all tools, Stripe built modular, agent-executable units where each unit encapsulates a specific task, its required tools, and the approach to solving it. This architectural choice, rather than a simple scaling trick, enabled Kai to reliably interact with hundreds of internal systems in an organized and maintainable way.
Further Reading
- How Stripe Built Their Knowledge AI Platform on Deep Agents - LangChain
- Meet Stripe's Knowledge AI Platform - Stripe Engineering
- Deep Agents vs LangChain vs LangGraph - LangChain
- Deep Agents Overview - LangChain Documentation
