Google Agent Executor Scales to Billions of Concurrent Agents: Open-Source AX Treats Sessions as Suspendable Actors
Production AI agents have a Kubernetes problem. The platform was built for stateless microservices and predictable batch jobs, not for programs that pause mid-task waiting on a model call, resume hours later, and accumulate state across thousands of tool invocations. Google’s answer is AX.
Agent Executor, or AX, is Google’s open-source runtime for distributed agent workloads at production scale. Announced via a Google Cloud blog post on May 20, 2026, the project is gaining renewed developer attention after the team revealed its full scaling ambitions last week: a compute runtime designed to handle billions of concurrent agent sessions per cluster, fitting 10-20x more sandboxes into existing hardware than standard Kubernetes configurations allow.
The Actor Model
The core abstraction: every agent, harness, skill, tool, or sandbox is an actor — a lightweight isolated unit that can be suspended, checkpointed, and resumed in sub-second time. AX pairs actors with Agent Substrate, a companion open-source project (github.com/agent-substrate/substrate) that replaces the standard Kubernetes control plane with one designed for the chatter of millions of sub-second tool calls that would overwhelm a standard scheduler.
Where Kubernetes is optimized for thousands of long-running stateless services, Agent Substrate manages hundreds of millions of registered agents by moving them onto and off ready compute in real time. The multiplexing ratio is the selling point: agents are idle most of the time, waiting on model responses or external tool results, so the same physical hardware can carry far more concurrent sessions than it would if each held a persistent process.
Five Runtime Capabilities
Durable execution: An event log and snapshot system that survives outages and human-in-the-loop interruptions. Long-running tasks resume exactly where they stopped, with no replay from scratch.
Secure isolation: Each actor runs in a sandbox — either microVM or gVisor — with hard enforcement boundaries. Malicious code inside one agent cannot reach adjacent sessions or cross tenant data boundaries.
Session consistency: A single-writer architecture prevents race conditions when multiple distributed components update shared agent state simultaneously.
Connection recovery: Clients can disconnect and reconnect to running agents. Missed responses backfill from the event log from the last sequence number the client acknowledged.
Trajectory branching: Checkpoints let developers fork an agent’s decision tree at any point, running different paths without losing accumulated context or state. This enables A/B testing on agent reasoning in production and is essential for RL evaluation loops over agentic trajectories.
Harness Ecosystem
Internally, Google routes Antigravity 2.0 and its Deep Research agent through the AX stack. Externally, the runtime is harness-agnostic: teams implement the HarnessService interface to bring their own execution harness, or use frameworks from the Agents Development Kit (ADK) or compatible agent protocols. The system federates across Google-managed agents, on-premises infrastructure, and custom deployments without lock-in to any model or cloud provider.
The roadmap includes: tool call approvals from harnesses, forking from event log snapshots, trajectory exposition for audit and evaluation, and improvements to resumption protocols. A significant architectural redesign is underway; external pull requests are paused while the core stabilizes, but issues and feature requests are open.
Project Status
GitHub: github.com/google/ax. Stars: approximately 1,973. License: Apache 2.0. Status: preview. The repo was created March 30, 2026; the public announcement came in May; architectural redesign began in September.
Why This Layer Matters
AX sits below orchestration frameworks like LangGraph, ADK, or OpenAI’s Agents API — and above raw Kubernetes compute. Until AX, this layer didn’t formally exist as an open primitive. Teams either built it themselves or accepted fragile long-running agent processes that failed silently during outages.
The trajectory branching capability deserves attention beyond reliability use cases. For AI labs running evaluation loops over agentic decisions — including RL training over multi-step agent behavior — a runtime that can fork a checkpoint and run N variants of the same trajectory concurrently changes what’s tractable to evaluate. That’s not incidental to the scaling story; it’s central to it.