Stripe has just revealed how its company-wide AI agent works. One engineer built the first version in a single week, and now 83% of Stripe employees use it every week.

This agent is called Kai. In Stripe's own words, it is a coding agent built for non-engineers. You chat with it, and it generates reports, dashboards, and documents.


## Two Key Design Choices

**1. What to load**

Stripe has over 500 internal tools and more than 1,000 skills. If all of them were crammed into the model's context window, it would overwhelm the model entirely. So Kai's default state loads nothing.

The model first determines which skills are required for the user's request, and only those skills bring their associated tools into the context. It works in two steps: first select the required skills, then load the corresponding tools.

Some netizens compared this approach to lazy loading — only load what you need when you need it, instead of dumping everything in at once.

**2. Where the code runs**

Kai itself runs outside the sandbox. The sandbox is treated just like any other regular tool that Kai can call.

This means one critical thing: any code written by the model can never access the agent that generated it. The execution boundary is clean, and a huge portion of potential security risks are eliminated immediately.

## A Candid Disclosure of Its Limit

Stripe openly admitted one issue: when the number of active skills goes beyond around 150, output quality starts to drop.

The team is still working to solve this problem.

Some netizens pointed out that this candid disclosure is more valuable than the 83% adoption rate. After all, most teams only share positive results when promoting their AI agents, and no one is willing to talk about where their system breaks down.

Another netizen shared their own experience: they once configured over 200 skills in Claude Code, only to find that the frontmatter alone took up a huge amount of the context window, leading to a noticeable drop in output quality.

## Can One Engineer Really Maintain This?

Someone raised a sharp question: can a single engineer really maintain an AI agent for the entire company? Doesn't it require a much more robust internal review process?

The answer lies in Kai's architecture itself. Skills are maintained separately by over 100 different teams. Core skills are loaded by default, while additional skills are loaded in layers based on user profiles and configurations. The engineer only maintains the framework, not the skills themselves.

## Key Takeaways

The reference value of this case lies in this: it is not a display of power by piling up computing resources. Instead, the team worked through engineering constraints clearly before starting development. Dynamic loading, sandbox isolation, layered skills — these are all actionable practices that teams can adopt directly.

As for the 150-skill limit? That's a problem left for the next team to solve.