Stripe's internal agent Kai is used weekly by 83% of the company. Its team saw quality drop past 150 skills, so load skills by role
The most useful detail in Stripe's agent story is not the adoption number. It is the point where adding more skills made the agent worse.
What happened
LangChain published a case study on Kai, Stripe's company-wide productivity agent. It runs on Deep Agents, LangChain's open-source agent harness, and one engineer built the first version in a week. Kai connects to Stripe's data warehouse, Slack and Google Suite, and produces reports, dashboards and documents that evolve as the chat goes on.
After open preview it grew from 296 users to over 5,000 in about four weeks. Today 83% of Stripe uses it every week, across more than 60,000 sessions. Marketing (95%) and GTM teams (87%) use it more by percentage than engineering, mostly for deal preparation and pulling data from several internal sources.
Teams own their own skills, and the library has passed 1,000 skills from over 100 teams, with 500+ internal MCP tools. Kai loads skills by user profile, so sales ops gets a different set than finance, and the chosen skills decide which tools load. The team found frontier model quality degraded with more than 150 skills combined with their system prompt. A few foundational skills stay pinned no matter what. Code runs in a sandbox the agent calls as a tool, not inside the agent's own environment.
My take
Most of my clients will never have 1,000 skills. They will hit the same wall at a smaller size, because the failure is the same: everything gets stuffed into one agent's context.
What I take from Stripe for a 20 to 200 person business:
- Build one agent and give each role its own skill set. Sales sees CRM and proposal skills. Finance sees invoicing and reporting.
- Let the team that does the work own its skill. The person who runs onboarding writes the onboarding steps.
- Pin the rules that must always apply, like which CRM fields the agent may never overwrite.
- Count your skills per role and test quality as the list grows. Stripe's line was around 150, yours may be lower with a smaller model.
The adoption numbers came from non-engineers. That is the audience worth designing for.
More posts
- Open OCR models now read 1,000 pages for under $0.10, and better serving saved more than switching modelsNov 20, 2026
- Gemini 4 Neon costs $3 in and $15 out per million tokens and tops Zapier's AutomationBench at 56.3%, though it uses more output tokens per taskNov 20, 2026
- Gemini API's Tool Retrieval loads tool schemas only when the model asks for them, so agents with hundreds of tools send fewer input tokensNov 20, 2026
