Agent Executor: Google Open-Sources Distributed Agent Runtime

A distributed agent runtime designed for reliability, safety, customization, and efficiency. It orchestrates agentic loops and manages execution through event logging, communicating with local and remote actors.

Controllers, skills, tools, and agents run as isolated Actors, with native support for automatic recovery and resumption upon failure or interruption. It uses a Single-Writer architecture and an event log to ensure consistent state management and durable execution state.

Key features include the following:

• Customization: Enables configuring purpose-specific agents tailored to particular tools, skills, domains, and workflows
• Built-in Gemini: Includes a Gemini agent for prompt-based text generation (requires AI Studio or Vertex AI authentication)
• Scalability and flexibility: A model- and harness-agnostic structure supporting deployments from small to large scale
• Kubernetes-native: Designed to run on Agent Substrate, optimized for K8s environments
• Interoperability: Supports agentic protocols such as MCP, A2A, enabling integration with various tools and agents
• Observability: Provides an audit trail, observability hooks, and trajectory collection for the entire execution

Users can perform execution (exec), running a server (serve), forking from checkpoints (fork), and visualizing traces (trace) through a CLI tool called ax. It is currently in early preview and is provided under the Apache 2.0 license.