Google AX and the 4-Layer Agent Stack
Google shipped AX (Agent eXecutor) two weeks ago. ADK is the Agent Development Kit. Agent Engine is the managed runtime on Vertex AI. AX is the third piece: open-source, self-hosted. The three are complementary.
What AX actually does
A central controller coordinates skills, tools, and agents in isolated sandboxes. It maintains a single-writer event log (one process owns state writes, so the log stays consistent across crashes), recovers automatically after failures, lets clients reconnect and replay from the last sequence seen, and exposes checkpoints to branch an agentic trajectory at any point. Human-in-the-loop pauses are handled natively.
AX is harness-agnostic and model-agnostic: it works with LangChain, LangGraph, ADK, agents via A2A, and tools via MCP. The framing is a runtime standard, not a proprietary Google stack.
It deploys on Kubernetes via Agent Substrate, Google's new abstraction since Kubernetes was built for long-running services, not sub-second tool calls. The demo multiplexes 250 actor sessions onto 8 pods (agents are mostly idle).
Agent Engine, by contrast, is Google's managed runtime on Vertex (autoscaling, cold starts, sessions). AX is its open-source, self-hosted counterpart.
The agent stack is splitting into layers, the same way it did with Docker and Kubernetes: container runtime first, orchestrator second.
For agents, the layers I see today are: 1) the model, 2) the SDK (ADK, LangChain, CrewAI, AutoGen, OpenAI Agents SDK), 3) the harness (tool dispatch, retries, context management), 4) the runtime (orchestrates layer 3 at scale, managed via Agent Engine or self-hosted via AX, Temporal, Restate, etc.).
Most teams fuse layers 3 and 4 in their own code, and that is what breaks in production. An agent crashes mid-run, state evaporates, you replay the trajectory from scratch at full token cost. AX names layer 4 as a separate concern.
AX is at 1.4k stars, the README says "active early development," external PRs are paused. Category announcement, not a production tool. Temporal has done durable execution since 2019. Restate is younger but already in production at several teams. Whether agent-native primitives (A2A coupling, checkpoint branching) justify a new runtime, or whether existing ones absorb them, is the open question.