AX is a new open-source declarative control plane for running stateful agent workloads at scale, built on Google's Agent Substrate.
AX launched on September 21, 2026, as an open-source declarative control plane for running agentic tasks at scale. Built on Google's Agent Substrate compute runtime, AX targets a gap traditional infrastructure struggles to serve: agents that hold state, burn compute in bursts, call external APIs, and idle while waiting for model responses or human approval.
Why agents need different infrastructure
Containers and serverless functions excel at stateless request-response work. Agents are different. They maintain conversation state, spawn sub-tasks, wait on human approvals, and retry tool calls when models hallucinate parameters. Running a fleet of agents on conventional Kubernetes jobs without purpose-built primitives leads to either over-provisioning or brittle custom orchestration.
AX abstracts agent execution into four core primitives: Task, Workspace, Gateway, and Model.
The four primitives
Task runs as an isolated sandbox with CPU and memory limits, cheap to create and discard. Each agent session maps cleanly to a Task lifecycle.
Workspace handles Git repository and MCP server setup before tasks start. This separates environment preparation from execution, which matters when agents need consistent tool access across retries.
Gateway enforces network policies and credential injection. After a month of sandbox escape headlines, network policy as a first-class primitive is not optional — it is the minimum viable security layer.
Model configuration lives in a single apply-able unit alongside the other primitives, so developers can version agent infrastructure the way they version application manifests.
Built on Agent Substrate
Google's Agent Substrate runtime provides the execution foundation. AX positions itself as the declarative layer above it — similar to how Terraform sits above cloud APIs. The project is open source and available on GitHub.
For teams already invested in Google Cloud's AI infrastructure, AX offers a path to scale agent workloads without rebuilding orchestration for every use case. For teams on other clouds, the primitive model is still worth studying as a reference architecture.
When to consider AX
Consider AX if you are moving from single-agent demos to fleets — customer support triage across time zones, code review bots on every pull request, or research agents polling external data sources. The declarative model pays off when agent count reaches dozens or hundreds and manual provisioning breaks down.
Skip AX if you are still validating product-market fit with one agent and one workflow. The complexity of a control plane only earns its keep at scale.
The open-source angle
Releasing AX as open source continues a trend in September 2026: infrastructure for agents is being commoditized quickly. WSO2 Agent Manager GA'd the same day for governance. ByteDance and Tsinghua open-sourced DAPO for training. Alibaba open-sourced Qwen-Image-2.1 for multimodal workloads.
Developers benefit from the competition. The winners will be teams that pick the right layer to differentiate — usually data, workflows, and UX — while leaning on shared runtimes for everything else.
Comments
Loading comments…