Why AI Agents Break in Production | Alex Salazar (CEO of Arcade) on MCP, Permissions & Enterprise AI

AI agents are rapidly moving from demos into production, but intelligence isn't the biggest challenge. Trust is.

In this episode, Kim Isenberg and Peter Thum sit down with Alex Salazar, co-founder and CEO of Arcade and former VP of Product at Okta, to explore why permissions, governance, and identity are becoming the critical infrastructure for enterprise AI.

Alex explains why today's agents often fail in production, why giving an AI agent unrestricted access is dangerous, how MCP differs from traditional APIs, and why guardrails alone aren't enough to secure autonomous systems.

The conversation also covers how Fortune 500 companies are deploying AI agents today, why human–agent collaboration is outperforming full autonomy, and what still needs to be solved before AI agents can safely operate across enterprise software.

Topics covered:

• Why AI agents fail in production
• Authentication vs. authorization
• MCP vs. APIs
• Enterprise security and governance
• Least-privilege access
• Human + AI workflows
• Agent hallucinations and risk management
• Fortune 500 AI deployments
• The future of enterprise AI infrastructure

Guest: Alex Salazar
CEO & Co-Founder, Arcade

Subscribe for more conversations with the builders shaping the future of AI, enterprise software, robotics, infrastructure, and frontier technology.