Google open-sources AX agent runtime on Kubernetes

Kubernetes was engineered for long-running stateless microservices, and shoehorning idle-heavy, tool-calling AI agents into traditional Pods has created a massive infrastructure mismatch that AX directly attacks.

• –Sub-second suspend and resume tackles the primary economic barrier of running autonomous agents by avoiding compute waste while waiting on external model calls or asynchronous tool loops.
• –Pairing gVisor-level sandboxing with strict Gateway egress policies delivers the defense-in-depth isolation enterprises demand before running untrusted agent code in production.
• –Offloading high-churn task lifecycle state to Redis Streams demonstrates realistic operational engineering, preventing control-plane etcd exhaustion under heavy agent workloads.
• –Broad developer adoption will depend heavily on frictionless interoperability with higher-level agent frameworks and the Model Context Protocol (MCP) rather than requiring teams to rebuild agent runtimes from scratch.