Google has announced a new open source project called AX, short for Agent Executor. The project focuses on managing and executing complex AI agent environments that run across multiple systems and perform long-running tasks.
According to Google, AX is intended as a distributed agent runtime. The platform is designed to coordinate agent workflows, log executions, and facilitate communication between local and remote components. The emphasis is on reliability and recovery capabilities when processes fail or are interrupted.
AX is currently still in an early development phase. Google warns that key parts of the architecture and runtime are still subject to change, meaning future versions will likely not remain compatible with earlier implementations.
A key feature of AX is support for so-called resumption. This allows AI processes to automatically resume after failures or interruptions. According to Google, this also applies to complex distributed environments where various agents, tools, and skills run as separate components.
To achieve this, the project uses, among other things, an event log where the execution status is stored. Additionally, Google mentions a single-writer architecture in which a single central controller remains responsible for consistent state management.
According to Google, AX is specifically designed for so-called long-running workflows: AI processes that can remain active for minutes, hours, or even days and must be able to handle human input, network interruptions, or system errors in the meantime. To this end, the platform supports features such as durable execution, which allows workflows to retain their status and continue after interruptions.
In addition, Google highlights features such as secure sandboxing to isolate agent components from one another, session consistency controls for distributed workflows, and connection recovery to maintain execution status during network outages. AX supports the selection and execution of skills, tools, and agents. All interactions go through a central controller, which Google says should simplify auditing and policy enforcement.
Testing alternative execution paths
AX also supports so-called trajectory branching. This allows developers to test alternative execution paths from saved checkpoints without losing previous context. This should make it easier to debug and optimize complex agent workflows.
Google also states that AX supports various deployment models. Organizations can create combinations of on-premises environments, Google-managed agents, and their own custom agents. The company also mentions support for the Agent2Agent (A2A) protocol.
According to InfoWorld, Google is positioning AX as an infrastructure layer for a new generation of AI agents. The company states that the market is shifting from relatively simple assistants to autonomous systems that remain active for extended periods and run distributed across multiple services and environments. Consequently, traditional monolithic agent architectures are expected to give way to distributed setups in which tools, skills, and agents function as isolated actors.
Although AX is compute-agnostic, Google says the project was primarily designed with Kubernetes environments in mind. The platform is intended to support both small and large-scale deployments. Google says it is intentionally releasing the project early to gather feedback from the community before a stable release is released. Developers can currently install AX via a Go package from GitHub.