Uploaded November 2025 | Updated September 2026, 18 hours ago

---

WorkOS’s Tobin South shares why small AI teams are selling to enterprises faster—and how to harden AI apps that started as “vibe‑coded” projects. He covers MCP/A2A realities, retrofitting security (e.g., client ID metadata), the gap between specs and implementations, and why standardized interfaces will unlock much stronger tool use.

CHAPTERS:
0:04 - Intro: leading AI agents at WorkOS
0:25 - Everyone’s building AI; complexity and security concerns
0:59 - Small teams vs enterprise features; vibe‑coded apps
1:55 - AI non‑determinism raises the bar for enterprise features
2:32 - New protocols: MCP and A2A; security often an afterthought
3:34 - Making MCP secure: client ID metadata; lock down registration
4:32 - Specs vs reality; implement and reverse engineer
5:45 - Standard interfaces → better tool use (RL on tool calling)
6:16 - Code skills complement MCP; MCP for SaaS interfaces