Moving from LLM Gateways to a Single Agent Origin

Moving from LLM Gateways to a Single Agent Origin ​AI gateways were designed for LLMs, but agents demand more – executing actions like trip planning or web searches, not just autocompleting text. Block/goose, built with LLMs and Model Context Protocol (MCP), delivers flexible, action-driven results. Envoy AI Gateway, co-founded by Bloomberg and Tetrate, evolved from LLM abstraction to power Tetrate Agent Router Service (TARS), now a key goose LLM provider. ​Since June 2025, MCP’s growth has integrated SaaS tools like GitHub, Postman, Context7, PGA, and Kiwi. ​This talk reveals how we designed the gateway’s agent layer and tested it with Goose recipes for robust end-to-end workflows. We’ll cover tricky technical areas around authorization, as well what sort of features a gateway provides, to help make your agents more robust. ​Key Takeaways: How to use a single origin for LLM and MCP calls, and what that gives you ​Reuse multi-step workflows with Goose recipes. ​Recent shifts in the agentic landscape and how Agent Client Protocol (ACP) changes things ​ ​Adrian Cole ​Principal Engineer at Tetrate ​Adrian is a principal engineer at Tetrate. He’s been a routine contributor to open source. Lately, he spends time on Agent Client Protocol (ACP), Envoy AI Gateway and Goose. His past work includes OpenTelemetry, wazero, Zipkin and jclouds. ​About the host / Chris Mazur Helping to connect the Software Development Community in the Southern Hemisphere since 2011. Software & DevOps Engineer, CNCF Ambassador. Building communities, one Mate 🧉 ( or a coffee ☕ ;) at a time! https://mate.dev/chris