Quick answer
Google Agent Executor—AX—is an open-source, self-hosted distributed runtime for agent execution and resumption. It is designed to run agents, harnesses, skills, tools, and sandboxes as isolated actors while preserving execution state in an event log. Google also describes snapshots, single-writer session consistency, reconnect backfill, and trajectory branching.
AX is not an agent framework and not a managed Google Cloud service. As of August 12, 2026, Google’s launch post calls it preview, while the repository warns that it is in active early development and may introduce major breaking changes before a stable release. That lifecycle boundary should control every adoption decision.
Where AX sits in an agent stack
| Layer | AX role | Still owned elsewhere |
|---|---|---|
| Agent or harness | Executes a registered or built-in harness | Prompts, models, planning, and business behavior |
| Runtime | Coordinates conversations, executions, streams, and recovery | Acceptance criteria and failure policy |
| State | Appends execution events and supports resumption | Retention, migration, backup, and deletion policy |
| Compute | Provisions isolated actors on a compatible platform | Cluster, capacity, images, network, and cost |
| Tools and skills | Makes them available to a supporting harness | Authentication, authorization, safety, and correctness |
| Operations | Exposes a runtime surface and optional telemetry configuration | SLOs, alerts, incident response, and rollback |
This separation matters. Durable execution can preserve a bad plan just as reliably as a good one. Isolation can narrow blast radius without proving tenant separation. A resumable stream can recover delivery without making a repeated external side effect safe.
The documented runtime primitives
Google’s launch announcement describes five core behaviors:
- Durable execution: event logs and snapshots support recovery after outages or human-in-the-loop pauses.
- Secure isolation: actors can run in separate sandboxed environments.
- Session consistency: a single-writer architecture reduces conflicting state updates.
- Connection recovery: a reconnecting client can receive responses after its last observed sequence.
- Trajectory branching: checkpoints can seed alternative execution paths without discarding prior context.
The current repository adds an important implementation frame. AX is compute-agnostic but Kubernetes-oriented, supports a local SQLite event log configuration, documents PostgreSQL for its Kubernetes path, and provides a gRPC server mode. The repository also says the preferred production deployment path uses Agent Substrate, while its deployment guide labels that path experimental and incomplete.
For recovery semantics and acceptance tests, read the event log and snapshots guide. For distributed client behavior, use the session consistency and connection recovery guide.
What AX is not
Do not evaluate AX as though it were a complete application platform:
- It is not a managed service; your team operates the runtime and its dependencies.
- It is not an agentic framework; a harness or agent implementation still defines behavior.
- It does not make a model, tool, MCP server, or skill trustworthy.
- It does not grant a caller authority to perform an action.
- It does not turn a Preview interface into a stable compatibility promise.
- It does not provide evidence that your workload meets a production SLO.
The harness, A2A, MCP, and skills guide explains the compatibility boundary. The security and multitenancy guide separates runtime isolation from application authorization.
A safe evaluation sequence
- Pin a repository revision, Go modules, container images, configuration schema, protocol definitions, and any Agent Substrate dependency.
- Start with a reversible task and a test identity whose permissions are intentionally narrow.
- Record an execution, interrupt it at controlled points, and verify resume behavior from both event and compute failures.
- Test duplicate prevention around every state-changing tool call.
- Disconnect and reconnect clients; confirm sequence ordering, backfill, cancellation, and terminal-state behavior.
- Run hostile cross-session and cross-tenant tests for filesystem, network, secrets, event-log, snapshot, and tool access.
- Rebuild from backup, roll back one version, and prove that stored state remains readable or has a migration path.
- Define an exit plan before expanding the cohort.
Use the Agent Executor readiness checklist to record the evidence. If Kubernetes is in scope, continue with the AX and Agent Substrate deployment guide.
Frequently asked questions
What is Google Agent Executor?
AX is Google’s open-source distributed harness runtime for executing and resuming agents and related actors in isolated environments.
Is AX a managed Google Cloud service?
No. The official repository says AX is self-hosted and not a managed service.
Is AX production ready?
The public project is Preview and in active early development. Production readiness must be proven for a pinned version and bounded workload; it cannot be inferred from the announcement.
Official sources
- Google Cloud: Agent Executor announcement
- Google: AX repository
- Google: AX Kubernetes deployment guide
Source check: August 12, 2026; repository revision 2bcc1637b3c106c16963c5c5464aedb46c6da031. Recheck the repository warning, README, protocol, manifests, dependencies, license, and release state before use.