Google's Agent Substrate Deep Dive: 250 Stateful Agents on 8 Pods, Plus Egress Gaps

ghumare64 · x · 2026-09-27

Developer Rohit Ghumare published a three-part teardown of Agent Substrate, the open-source agent runtime stack Google Cloud is building (README demo: 250 stateful agents on 8 pods).

Part 1 — Warm pool runtime: each agent is an actor living in a snapshot, borrowing a warm worker pod only while working. Size the pool on peak concurrency: 400 agents at 9% busy ≈ 40 workers. Apache-2.0, sandboxed via gVisor/Kata, available on GKE for non-production since Sept 15.

Part 2 — Google AX (Agent Executor): declarative Task/Workspace/Gateway/Model objects. A resumed task is a fresh process over a restored /workspace — in-memory state is gone, so write agents as if every suspend were a crash with 10 seconds of warning. A Sept 20 commit rebuilt AX into a general-purpose orchestration layer (662 points on HN).

Part 3 — Egress gaps: the gateway ignores claimed Host/SNI and trusts only the kernel-seen IP; the control plane authenticates but has no authorization yet, so any authenticated caller can rewrite any agent's egress policy. A CNCF donation is in progress.

Original post →

More from coding & agent

coding & agent channel →