Google introduces AX, an open runtime for running AI agents at scale

Google has published documentation for AX, an open framework for running and orchestrating autonomous AI agents, under the project name agentexecutor.io. The project describes agent workloads as sandboxed processes, called actors, that carry out a declared goal, such as ensuring a programming-language toolchain is installed and built from source, inside an isolated environment with set resource limits.

According to the documentation, AX organizes agent infrastructure around four elements. Sandboxes run untrusted agent code under CPU and memory limits and are meant to be cheap to create, suspend and throw away. Workspaces are set up either by listing the Git repositories, MCP servers (Model Context Protocol servers, which connect agents to external tools) and skills an agent needs, or simply by describing the goal in plain terms. Network policies restrict an agent’s traffic to an explicit allowlist of hosts and ports and can inject credentials into outgoing requests. A fourth element centralizes model configuration, parameters and secrets, so a developer can rotate a key or pin a new model version in one place.

AX’s compute layer is built around what the documentation calls fast, stateful actor lifecycles. Agents that are idle, for instance while waiting on a model response, a tool call or human approval, are checkpointed and suspended in under a second with no cold-start delay when resumed. Multiple tasks share worker resources, which the project says turns idle waiting time into spare compute capacity so that cost is tied to when an agent is actively computing rather than to how long it stays alive.

Generative setup and intended uses

AX also integrates generative AI into its own setup process. Rather than scripting an environment by hand, a developer can describe what a ready workspace should look like in plain English, and AX hands that description to an agent on first boot to install toolchains and verify dependencies. The documentation lists interactive coding agents, long-running agent servers, Jupyter notebooks, headless browser testing and custom tool runtimes as intended uses, alongside spinning up large numbers of reproducible sandboxes to collect training trajectories, run reinforcement learning loops, and evaluate agents at scale.

According to the project documentation, AX originated inside Google, where researchers working on agentic runtime systems concluded that agent workloads form a distinct computing pattern: stateful, bursty processes that compute intensely for short periods and then wait on models, tools or people. The project says traditional orchestrators, built for microservices or predictable batch jobs, become costly when they must keep idle sandboxes running and lack native support for fast suspension and resumption. AX draws on agentic runtime research from Google DeepMind and depends on a separate, previously built Google project called Agent Substrate for underlying isolation, scheduling and resumption components, layering agent-specific abstractions on top.

Google has not published a license, pricing or a general-availability timeline for AX in the material reviewed. Developers can currently review the project through its Hacker News post and the agentexecutor.io site, where further technical documentation is expected to appear as the project develops.