Google has open-sourced AX, short for Agent Executor, a distributed runtime for running agent workloads at scale. Announced alongside Google Cloud's agent stack, AX takes a declarative YAML task and handles the plumbing around it: it sandboxes the task, wires up its workspace, fences its network, and schedules execution across a cluster. Google says the design targets billions of tasks per cluster, and the project hit 485 points and 213 comments on Hacker News within a day of release.
The runtime targets problems that appear once agents run for hours rather than seconds. An event log plus snapshotting gives every actor durable execution, so a crashed agent resumes instead of restarting from scratch. Secure-by-design sandboxes contain harmful side effects when models generate code or handle multiple tenants. A single-writer architecture keeps shared session state consistent, clients can reconnect after network drops and get backfilled responses, and checkpoints let teams branch an agent trajectory to test alternative paths without losing context.
AX is positioned as harness-agnostic and portable. It can mix Google's Antigravity harness, Gemini frontier agents, custom LangChain or ADK builds and any A2A agents, and run on a team's own data plane so workflows stay off a vendor's compute. Google also announced Agent Substrate with the GKE team, an open-source Kubernetes layer for moving agents onto and off ready compute. For founders building agent products, the takeaway is straightforward: scheduling, recovery and isolation are increasingly infrastructure someone else already maintains.