Google has open sourced their internal agent orchestrator.
retained text
Google shipped AX, their open source agent runtime built on Agent Substrate. The crazy part for me is how they handle task state. Instead of putting millions of short lived agent tasks into Kubernetes and etcd, AX stores the state in Redis and reconciles directly with Agent Substrate. The idea is etcd works well for Kubernetes control plane state, but millions of short lived agent tasks create a very different kind of churn. It’s basically an admission that we’ve been running agent state through infrastructure designed for a different problem: long lived services. Different scale obviously, but it’s the same question I run into with my own setup, just kinda smaller like where does the state live when more than one agent is touching the same thing at once? Google is like infrastructure built for a fleet given for me it is Tutti keeping the live state visible for the handful of agents I’m running day to day. Same core problem showing up at both ends of the spectrum. you guys have dug into the AX architecture yet ( I went on a rabbit hole for this )
document fetched 2026-10-03 00:18 · canonical https://reddit.com/r/AI_Agents/comments/1wvzaep/google_has_open_sourced_their_internal_agent