From Demos to Deployment | Alex Salazar (CEO, Arcade.dev)

Listen to full episode:

Summary: This episode explores the hard part of bringing AI agents into production: not the demo, but the security, authorization, and accuracy problems that show up the moment an agent needs to do real work. Anant and Ed are joined by Alex Salazar, cofounder of Arcade.dev, who explains why the last mile of agentic success is really a control plane problem. Alex also shares how Arcade.dev is trying to solve the gap between agent capabilities and enterprise readiness by combining OAuth-style authorization, tool governance, and accuracy-focused developer primitives.

Chapters:

00:00 — Browser wars & why the browser is becoming an AI interface

02:12 — Browser-based agents: trust and prompt injection concerns

07:08 — Anthropic, MCP, and code-first agent execution

10:19 — Why enterprise agents need a control plane, not just a model

12:44 — Arcade.dev's pitch: MCP runtime, authorization, governance

14:55 — Permissions problem: agent failure patterns & the right permissions model

21:27 — Most enterprises stuck at level one; MCP is early but pressure to ship is here

25:21 — What developers need: system access, and why accuracy beats speed

28:54 — Enterprise agent architecture vs. web apps; new AI/ML-oriented teams

35:12 — Trust nobody: agent, application, workflow engine, and the CISO's hard role

39:53 — Accuracy is a slog: evals, tool definitions, retries, and Arcade.dev's pivots

45:31 — Drop-in security, developer resistance, and why demos aren't enough

48:13 — Closing: interface direction, AI-quantum convergence, thanks

Sound Bites:

“You can't trust the agent. You can't trust the application.”

“The primary performance metric of an agent is not speed, it's accuracy.”

“There is no silver bullet to accuracy. It is just a slog.”