Agent Executor (AX)
A distributed runtime for autonomous agents with resumable isolation, suspension, and recovery.
Evidence shows: README describes isolated execution, event log, and recovery, but lacks specific least-privilege implementation details. User confirmation mechanism not mentioned. Data flow transparency has event log but no data flow direction. Sensitive data handling only mentions API key env vars, no storage/transmission security. Dependency security has go.mod but no vulnerability scanning or version pinning. External effects include MCP tools and filesystem ops, but no permission controls. Rollback has recovery but not explicit rollback. Source attribution clear with Google org and Apache 2.0. Deductions: lack of implementation details and user control mechanisms.
Evidence shows: README and code structure consistent, but no test results. Dependency availability has go.mod but not verified. Failure messages have recovery but no error details. Deductions: no test evidence and error handling details.
Evidence shows: README identifies target users (developers/researchers) and use cases (distributed agent runtime). Capability boundaries have 'What AX is NOT' section. Trigger precision has CLI options but no precise trigger conditions. Environment fit has Kubernetes support. Deductions: trigger precision and boundary descriptions not detailed enough.
Evidence shows: README structure clear, install notes detailed, naming stable (AX), examples and FAQ (partial), known limitations have 'What AX is NOT', license Apache 2.0, versioning changelog not provided, maintenance responsibility by Google team. Deductions: missing changelog and incomplete FAQ.
Evidence shows: Output usability has CLI and examples, marginal value has distributed runtime and recovery, cost-benefit lacks performance/resource data. Deductions: cost-benefit lacks data support.
Evidence shows: README claims partially supported by code, but no cross-validation. Fact/inference separation unclear. Deductions: lack of tests and external verification.
- Project is in early development with major breaking changes and external contributions paused.
- No security audit or vulnerability scanning evidence; dependency security needs further verification.
- User confirmation mechanism absent, may affect security and compliance.
What does this agent do, and when should you use it?
AX (Agent Executor) is an open-source distributed agent runtime built by Google. It dynamically provisions isolated environments from suspendable/resumable images to execute harnesses and agents. Its core emphasis is reliability, featuring an event log for durable state, a single-writer architecture for consistency, and compute-layer actor resumption. While compute-agnostic, AX is optimized for Kubernetes, especially with Agent Substrate for production deployments. It includes a built-in Antigravity harness supporting Google AI Studio and Vertex AI, and supports custom harnesses, MCP tools, skills, and environment customization.
AX provides a CLI and server component. The CLI starts a local built-in AX server and launches the Antigravity harness (default) to process user input. Execution state is stored in an event log (SQLite by default), enabling resumption of conversations from any point, even after failures. In remote mode, the CLI connects to a gRPC server (via 'ax serve'), which coordinates actor execution. AX can dynamically schedule isolated actor instances on Kubernetes via Agent Substrate, allowing suspension and resumption for efficiency. Users can pass per-request agent configuration via '--agent-config' or '--agent-config-file'. The harness supports Agent Skills and MCP tools.
- Developers needing durable execution for long-running or bursty agentic workloads.
- Teams deploying agents on Kubernetes that require high density and isolation.
- Agent applications needing fine-grained control over tools and skills with isolated execution.
- Operators managing multi-tenant agent deployments that require recovery from failures.
- Users building agents with Gemini or Vertex AI who want a robust runtime with logging and resumption.
What are this agent's strengths and limitations?
- Strong resumption and suspend/resume capabilities for long-running or intermittent agents.
- Framework-agnostic, allowing custom harnesses and MCP tool integration.
- Native Kubernetes support via Agent Substrate for scalable deployments.
- Event-logging and single-writer design ensure consistency and auditability.
- Early development stage; APIs may break and external PRs are paused.
- Primarily tuned for Google's Gemini and Vertex AI; other models require custom harness work.
- Requires Go and Kubernetes knowledge for deployment; not a managed service.
- Planned features like elicitation and FilesystemService are not yet available.
How do you install or deploy this agent?
Install the ax CLI using Go: go install github.com/google/ax/cmd/ax@latest. Verify with: ax --help. Requires Go. Set Gemini API key (export GEMINI_API_KEY="your-api-key") or configure Vertex AI auth (gcloud auth application-default login and set GOOGLE_CLOUD_PROJECT/LOCATION).
How do you use this agent?
Run a simple example: ax --input "Can you list this directory?". To use a remote server: ax --input "..." --server localhost:8494. Resume a conversation: ax --conversation <id> --input "...". Resume an incomplete execution: ax --conversation <id> --resume. Start the server: ax serve --config ax.yaml (sample config includes server address and SQLite event log).