Stripe Built Its Kai Productivity Agent in One Week Using Deep Agents
- Cargo
- Machine Learning
- Shipped
- Shipper
- John Brown
- Readers
- 680
Kai connects Stripe employees to internal data, tools and workflows; adoption reached 83% of employees each week, while skill selection remains a scaling challenge.

Stripe’s Kai shows how a shared agent foundation can help a company bring AI to employees whose work doesn’t happen in a terminal. Built in one week by a single engineer, Kai gives staff a session based workspace for research, analysis and document creation. Its adoption also exposed the next set of engineering problems: selecting relevant skills from a large catalog, adapting behavior to different users, and managing access as usage expands.
From agent experiments to a company-wide assistant
Stripe’s AI platform team supports the infrastructure behind the company’s internal AI applications. As interest in agents grew, engineering groups were building orchestration layers on top of Stripe’s Ruby and Java systems. Connecting those efforts to internal services was manageable; making agents dependable across complex tasks required more work on performance, evaluation and consistency.
The launch of Claude Code in late 2024 brought agent building to more employees, but non-engineers faced practical barriers around terminals, data access and security. Sharadh Krishnamurthy, engineering manager for the Agent Foundation team, and Anupam Upadhyay, staff software engineer, saw an opportunity to provide an assistant designed around employees’ existing work. That became Stripe’s Knowledge AI Platform, or Kai.
Kai is available to every Stripe employee through a session based interface. A user can ask questions, analyze trends, brainstorm or draft material, with connections to Stripe’s internal data warehouse, Slack and Google Suite. The conversation can produce reports, dashboards and documents that sit alongside the chat and change as the session develops. The product is designed to give non-engineers some of the iterative workflow associated with coding agents.
Context comes from a library of more than 1,000 skills contributed by over 100 teams. These skills encode knowledge of Stripe’s internal systems, data sources and ways of working, so employees don’t need to reintroduce their role or company context for each task.
Four layers separate the agent foundation from the product
Kai runs on Deep Agents, LangChain’s open source agent harness. The foundation supplies common agent capabilities such as model request handling, tool calling, middleware composition, streaming and state management. Stripe’s team could build its internal integrations and workflows on top of those primitives instead of creating and hardening the whole agent runtime itself.
The architecture has four layers. Deep Agents is the base; a Stripe specific harness connects it to the company’s security model, infrastructure and services; a configuration layer lets teams define agents with distinct skills and behaviors; and the Kai interface connects employees to those capabilities. This arrangement lets teams change an agent’s configuration without rebuilding the underlying harness.
Krishnamurthy described the division this way: “The Deep Agents layer solves all the non-Stripey problems, so that we can focus on solving the Stripey agent problems.” He also pointed to reusable middleware and composable skills as a way to retain flexibility while providing structure.
Persistent files, isolated execution and long sessions
Several middleware components helped make Kai suitable for production. Because Kai runs as a cloud service, Stripe built a virtual filesystem backed by S3. Agents can read, write and refer to files as context across turns. Before code runs in a sandbox, relevant files are copied into it; afterward, new and changed files are synced back to the virtual filesystem. This “sync in / sync out” pattern gives the agent a consistent file environment throughout a session.
For analytics and document processing, Kai can call a sandboxed execution tool. It uses Python to query data and create charts, and it can process formats such as PDFs, presentations and structured documents. The agent itself runs outside the sandbox and invokes it as needed. Keeping generated code execution in a separate environment establishes a clear boundary between the agent and the code it runs.
Summarization middleware helps Kai handle long, multi turn sessions without carrying an ever larger conversation into each request. Stripe tunes the summarization threshold, the model that produces summaries and the output size. Those choices balance retained context against model limits and cost. Since users often return to sessions intermittently, avoiding large context cache misses also matters to operating expense.
Skills provide access to hundreds of tools
Stripe uses a federated model for its skills: individual teams maintain the skills for their own workflows. Kai starts with foundational skills for navigating Stripe, then loads additional skills based on the user’s profile and the configured agent. Someone in sales operations may receive a different set from someone in finance, and individuals can add skills on top of the defaults for their function.
The system has more than 500 internal MCP tools, so placing every tool in the model’s context at once would be impractical. Instead, skill selection controls which tools are loaded. Agent Skills’ allowedTools lists support this two step process: the model first chooses relevant skills, then the system loads the tools associated with them. Upadhyay said the team relies on the model’s context to select skills and, through that selection, the corresponding tools. A set of foundational skills stays pinned to preserve consistent company context and behavior.
Selection quality has limits. Stripe found that with a 1,024 character frontmatter limit, frontier models’ performance degraded when more than 150 skills were included alongside the system prompt. The full library now exceeds 1,000 skills, so deciding which ones to expose remains active work.
A week to build, then rapid adoption
Stripe had spent more than a decade building internal tools and deployment support around Ruby and Java. Supporting a Python based stack meant investing in service scaffolding for another language. Kai provided an immediate demonstration of what that investment could enable: Upadhyay built its initial version in one week, using Deep Agents’ existing infrastructure. Chrissie, head of AI Platform, said the project “completely solidified people’s belief that Deep Agents is the way to go.”
After Kai entered open preview, it reached its quarterly adoption target in a week. Its user count grew more than 16 times, from 296 to over 5,000, in roughly four weeks. The early audience included sellers, finance analysts and business operations staff who had been encouraged to use AI but found existing tools difficult to approach.
At the time described, 83% of Stripe employees used Kai each week, across more than 60,000 sessions. Adoption by percentage was higher in Marketing, at 95%, and go to market teams, at 87%, than among engineers. Employees use it to prepare for deals by combining information from internal sources and turning it into usable materials; finance teams use it for analysis and dashboard creation. New hires also use Kai during onboarding, while engineers use it to query internal systems and examine traces and usage data for gaps in skill coverage.
Employee reactions included claims that Kai could save salespeople tens of thousands of hours and that it marked a before and after in one person’s Stripe career. Those comments point to perceived value, while the usage figures show how quickly the tool reached employees beyond engineering.
Scaling selection and adapting to people
Stripe’s next work follows from Kai’s reach. The team is developing a hybrid approach to skill selection. Pure model selection works at the current scale and, when the model has full context, has produced better results than retrieval augmented generation. But the team expects a retrieval or classifier step to filter the catalog before the model makes its final choice as the number of skills grows.
Security and compliance guardrails also need to keep pace with use across thousands of employees. At the same time, users want different interaction styles: some prefer planning and clarifying questions, while others want quick, direct responses for brainstorming. Stripe is assessing classifiers and AI based methods that could infer the right behavior from user context, consistent with Kai’s aim of avoiding manual settings.
Today, Kai sessions are for one user. The roadmap includes shared skills and artifacts, followed by collaborative sessions in which multiple employees can work on the same AI generated outputs. For enterprise teams adopting internal agents, Kai’s record suggests that fast construction is only the first test; sustaining usefulness will depend on how well the system routes people to the right expertise and supports work across team boundaries.
Kai is built on Deep Agents, LangChain’s open source agent harness. Technical details are available in the Deep Agents documentation.