Stripe Reveals Internal AI Agent Architecture; LangChain Launches Managed Deep Agents
hwchase17 · x · 2026-08-09
Stripe shared the architectural details of Kai, its company-wide AI agent. Built by a single engineer in one week, it is a coding assistant for non-engineers now used weekly by 83% of employees.
Key Architectural Designs of Kai:
- Dynamic On-Demand Loading: With 500+ internal tools and 1,000+ skills, nothing loads by default to prevent token overload. The model first selects the necessary skills based on the request, which then bring in their specific tools—a two-step process.
- Sandbox Isolation: The agent sits outside the sandbox and calls it as a tool, ensuring model-generated code can never reach the agent itself. Stripe notes that quality starts to degrade past 150 skills.
Additionally, LangChain founder Harrison Chase reshared the deep dive and announced Managed Deep Agents, discussing the evolution of agent orchestration to provide an easier way to build and deploy production agents.
More from coding & agent
- Developer Recreates Classic iBeer App as an MCP Experiment — Puzzleheaded_Mine392 · 2026-08-09
- Lilian Weng's Deep Dive: The Key to Agent Self-Improvement Lies in Harness Engineering — aigclink · 2026-08-09
- Adversarial Testing Catches Silent Prompt Injection Regression in Doc Assistant — OpeningBird6240 · 2026-08-09
- Why Is There No 'App Store' for Independent AI Agents Yet? — mgsz_ · 2026-08-09
- Vibe Coding Accelerates Shipping, But Outpaces User Validation — Ubunta · 2026-08-09
- Building Production AI Agents: Linear Team Shares Core Lessons — petergyang · 2026-08-09