Interview

Why AI agents break in production

Alex Salazar

CEO & Co-Founder, Arcade

68 minHosted by Kim Isenberg and Peter Thum

Watch on YouTube

About this conversation

AI agents can write code, send emails and act across enterprise systems. But the moment they move from demo to production, a harder question appears: what are they allowed to do — and who can prove they followed the rules? Alex Salazar, co-founder and CEO of Arcade and former VP of Product at Okta, explains to Kim Isenberg and Peter Thum why permissions, not model intelligence, are the real bottleneck for enterprise AI agents.

They discuss why giving an agent super-admin access is a non-starter, how MCP differs from an API, why governance cannot live inside the agent, what Arcade sees in enterprise deployments, and why human-plus-agent workflows are outperforming fully autonomous agents.

Topics

  • Why impressive agent demos break in production
  • Authentication vs. authorization for AI agents
  • The three common permission failures enterprises make
  • MCP vs. APIs, explained in plain English
  • Why guardrails and governance are not the same thing
  • Why Human + Agent is beating full autonomy today
  • Who owns the failure when an AI agent makes a harmful move

Chapters

  1. 0:00Why AI agents break in production
  2. 1:01Introduction: Alex Salazar, CEO of Arcade
  3. 2:07How Arcade discovered the agent permissions problem
  4. 4:39Authentication vs. authorization for AI agents
  5. 6:56What Arcade does for enterprise agents
  6. 8:23Why Arcade pivoted away from building its own agent
  7. 10:51Why agent permissions are so complicated
  8. 14:10The three biggest permission failures
  9. 19:59Can an agent safely request more privileges?
  10. 24:23What makes an enterprise AI agent trustworthy?
  11. 25:23MCP vs. APIs explained
  12. 28:02Are AI agents users, apps, or non-human identities?
  13. 31:47Why governance must live outside the agent
  14. 32:25Guardrails vs. access control
  15. 35:32How Arcade scales thousands of integrations
  16. 41:14What an MCP gateway actually does
  17. 43:40Granular permissions and chained tool calls
  18. 46:32What Fortune 500 agent deployments reveal
  19. 48:47Why Human + Agent beats full autonomy
  20. 50:12Arcade’s role in the MCP authorization standard
  21. 52:03Sensitive data, regulated industries, and accountability
  22. 55:37How AI will reshape careers and jobs
  23. 59:14What agent infrastructure still hasn’t solved
  24. 1:02:11How Arcade plans 24 months ahead
  25. 1:06:19The agent assumption the industry will abandon
  26. 1:06:58Where to follow Alex and Arcade

Kim Isenberg is co-founder and Editor-in-Chief of Superintelligence, an AI newsletter published every weekday.

Read the Newsletter

More interviews

All interviews