Google Open-Sources AX and Agent Substrate: Reinventing Kubernetes for Agents

Google engineer rakyll (Jérémie) open-sourced two agent infrastructure projects on Sept 20-21: the agentic orchestrator AX (google/ax) and the agent execution runtime Agent Substrate. Each gathered 2k stars at launch; rakyll frames the work as "reinventing Kubernetes for agentic workloads."

Confirmed

• AX is Google's open agentic orchestrator and runtime built around statefulness and fast resumption for massive long-running agent workloads. Developer arpitbhayani's first impression: it feels like "Kubernetes for agent execution."
• The compute runtime targets billions of agents per cluster, drivable by a single natural-language instruction (e.g., "set up a Python 3 environment"), with reproducibility, branching and forking.
• rakyll claims 10-20x agent sandbox density on existing deployments, plus easy-to-use agentic abstractions and generative platform capabilities.
• Agent Substrate is a secure-by-default agent execution runtime on Kubernetes/GKE. Per dlweekly, it runs millions of sandboxes at 10x standard container-runtime density, with 1,000+ dormant agents per host. rakyll says it is production-ready with fast community growth.
• Its core selling point is extremely fast suspend/resume — fast enough to suspend between individual tool calls — underpinning rakyll's thesis that "the agent loop is the new scheduling loop." jonathangrahl notes Substrate powers all of rakyll's earlier demos.
• Substrate ships an interactive installer to lower the onboarding barrier.

Unconfirmed

• All performance figures (10-20x sandbox density, billion-agent scale, 10x container density, 1,000+ dormant agents per host) are Google/rakyll claims pending third-party validation.

Why it matters

• Traditional Kubernetes assumes stateless services, a poor fit for long-lived stateful agents; Google open-sourcing this stack signals agent infrastructure as the cloud vendors' next battleground.
• Per-tool-call suspend/resume, if it holds up, dramatically raises agents-per-machine and cuts deployment costs for large agent fleets, offering a reference implementation for agent-native cloud platforms.

2026-09-20 ~ 2026-09-22 · 14 related posts

• [source] Google open-sources AX, a Kubernetes-style orchestrator built for agentic workloads — rakyll · 2026-09-20
• [source] Google's AX runtime promises 10-20x more agent sandboxes on existing setups — rakyll · 2026-09-20
• rakyll Is Building a Compute Runtime That Scales to Billions of Agents — rakyll · 2026-09-20
• Google building open-source agentic orchestrator for billions of agents, rethinking Kubernetes — gaganghotra_ · 2026-09-20
• [source] Agent Substrate open-sourced: agent runtime with 10x container density, sub-500ms resume — rakyll · 2026-09-21
• Agent Substrate launches: suspend/resume agents per tool call on Kubernetes — rakyll · 2026-09-21
• Substrate bets the agent loop is the new scheduling loop with fast suspend/resume — rakyll · 2026-09-21
• Google open-sources Agent Substrate on GKE: 10x container density, sub-500ms agent resume — dl_weekly · 2026-09-21
• Google reveals it's rebuilding Kubernetes for agentic workloads and an open-source orchestrator for billions of agents — gaganghotra_ · 2026-09-21
• Google open-sources Agent Substrate, a K8s agent runtime that suspends and resumes per tool call — blaizedsouza · 2026-09-21
• Google open-sources AX, an agentic orchestrator likened to Kubernetes for agent executions — arpit_bhayani · 2026-09-21

3 near-duplicate retellings: rakyll · jonathangrahl · Existing_Hat_1064