Google Rebuilds Its AX Agent Orchestrator Around Redis, Not Kubernetes

Google restructured AX, its open-source orchestrator for autonomous agent workloads, in a commit merged Sept. 20 that removes the project's Kubernetes-based task state entirely in favor of Redis Streams. The change, authored by Google engineer JBD under the GitHub handle rakyll, is tagged as release v0.3.0.

The commit message describes AX as "undergoing a significant redesign, moving away from a single CLI with an embedded Python harness toward a general-purpose orchestration layer for agentic tasks." It reorganizes the repository around three binaries: ax-server, which exposes a gRPC API for submitting task manifests; ax-controller, a horizontally scaled reconciler that consumes work from Redis Streams; and ax-task-runner, which executes tasks inside sandboxed workers.

Why Kubernetes custom resources had to go

Task state previously lived in Kubernetes custom resource definitions, an approach the project says strains etcd, Kubernetes's own storage layer, once task volume climbs into the millions. Moving that state to Redis is meant to let AX handle millions of short-lived tasks without hitting that ceiling. The rewrite also removes the legacy Python harness, an older client library and a SQL-based event log, replacing them with new manifest, networking and sandbox documentation.

What AX has claimed to solve since May

Google introduced AX as an open-source runtime on its own blog around May 25, and outlets including InfoWorld and Techzine covered it within a day of each other. The pitch, per Google's post as quoted by InfoWorld, centers on "durable execution, allowing workflows to resume after outages or human approvals, along with secure sandboxing for isolating agent components." The project frames four primitives: a Task that runs untrusted agent code in an isolated sandbox, a Workspace that pre-configures Git repositories and MCP servers, a Gateway that restricts outbound network traffic to an allowlist, and a Model that manages which LLM an agent calls and with which credentials.

AX component (v0.3.0)	Role
ax-server	gRPC API for submitting Task manifests
ax-controller	Reconciler that reads work from Redis Streams
ax-task-runner	Runs tasks inside sandboxed workers

The skepticism that greeted the launch

Advait Patel, a site reliability engineer at Broadcom, told InfoWorld that "durability, orchestration, and resumability are the real blockers for enterprise production agents," adding that frameworks like LangChain "fall apart in production once agents run for hours or days." Gaurav Dewan of Avasant was more cautious, saying accountability and policy enforcement for agents were "still evolving" and that enterprise buyers would need governance beyond runtime infrastructure alone.

Google's own documentation says AX remains in an early development phase and that "key parts of the architecture and runtime are still subject to change," a caveat the Sept. 20 restructuring makes literal. The project had 5,538 stars and 231 forks as of Sept. 21, under an Apache 2.0 license, and competes for attention with runtimes like AWS's AgentCore, which targets the same durable-agent problem from inside AWS's own cloud. AX's star count is smaller than the 2,428 a Cloudflare security skill gained in a single day on GitHub last month, a reminder that trending and adopted are not the same thing on that platform.

Google has not said when the three-binary architecture will be considered stable, or whether the Python harness it just removed will return in another form.