Episode 219: Agentgateway: Keep Infrastructure Out of Your AI Agents

The Linux Foundation Agentic AI Foundation (AAIF) recently accepted the agentgateway open source project as the fourth project alongside MCP (spec), AGENTS.md, and Goose. Conceptually, an agentgateway is similar to an API gateway with some specific optimizations: handling of request body, multiplexing, and advancements in identity/authorization mechanisms. Is an agentgateway critical for AI agentic governance and control in enterprises?

#Speakers

• Host: Mike Schwartz, Founder/CEO Gluu
• Guest: Christian Posta, Field CTO, Solo.io

#Homework

• agentgateway Github
• Enterprise MCP gateway security in practice with agentgateway
• What Is an Agent Gateway, and Why Does Agentic AI Need One?
• Linkedin PostCredential Vault Animation

#Takeaways

• You can’t govern what you can’t see. agentgateway first provides observability; authorization, quotas, tracing, and other controls follow.
• agentgateway is CEL-native, Cedar-compatible but not Cedar-integrated. CEL makes sense for gateway expressions and transformations. For HTTP, connect Cedar through ext_authz. For MCP, implement a Cedar-backed ExtMcp policy processor. The latter is particularly compelling because it gives Cedar access to the MCP arguments that agentgateway deliberately withholds from native request-time RBAC.
• agentgateway provides a way to use services that require credentials by vaulting them, on the theory that at least in the gateway, we can better protect them.
• An agent gateway is infrastructure, not a security panacea. Zero Trust still requires enforcement throughout the stack—not one giant trusted gateway.