The Pulse
Google Open-Sources Agent Executor for Long-Running Agent Workloads
Google has released Agent Executor, an open-source runtime for durable, secure and distributed agent execution, alongside Agent Substrate, a Kubernetes-based compute layer designed for large-scale agent workloads.
AI.info Team ·
“long-running agent workflows are fragile and incredibly hard to manage reliably and efficiently in production.”
Jaana Dogan, Software Engineer, and Ethan Bao, Engineering Director, Google Cloud
Google announced Agent Executor on May 20, 2026, as an open-source runtime standard for agent execution, resumption and distributed deployment. The project is designed for agents and related components that may run for hours or days, lose connectivity, wait for human input or need to recover from an interruption without starting over.
Agent Executor provides several capabilities for long-running workflows. Its durable-execution features use event logs and snapshotting to support resumption after outages or human-in-the-loop confirmations. Secure sandboxes isolate components such as agents, agent harnesses, tools and other workloads. A single-writer architecture helps multiple components update shared session state without corrupting it, while connection recovery allows clients to reconnect and receive responses from the last sequence they observed.
The runtime also supports trajectory branching. Checkpoints can be used to branch an agent’s decision or workflow path, allowing developers to test or evaluate alternate paths without losing the surrounding context or state.
Federating Different Agent Deployments
Google presents Agent Executor as a way to coordinate agents across different deployment models. Organizations may need to keep proprietary workflows on-premises for performance, compliance or control reasons, while also using managed or custom agents. The runtime is intended to connect those environments.
Google says Agent Executor can work with Google’s Antigravity agent harness, Google-built agents such as Deep Research, custom agents managed through the Managed Agents API, and purpose-built agents developed with frameworks and protocols including LangChain, LangGraph, the Agent Development Kit and Agent2Agent Protocol.
The project is also designed to let organizations retain control over their agents, models and compute. Google says developers can deploy agents on their own infrastructure without being tied to a particular provider’s model or compute environment. Agent Executor is harness-agnostic, supports industry-standard frameworks and protocols, and allows developers to run components such as MCPs, skills and other agents on their own data plane.
Agent Substrate Adds an Agent-Focused Compute Layer
Google introduced Agent Substrate alongside Agent Executor as a separate open-source project developed with the Google Kubernetes Engine team. Agent Substrate is intended to address the compute and control-plane demands created by large populations of agents that may spend long periods waiting for external inputs.
Google says agent workloads are scaling into the hundreds of millions and are increasingly long-running. Unlike conventional software, agents are nonlinear programs that may pause for human interaction, events or other triggers. Agent Substrate moves agents on and off ready compute capacity in real time, using Kubernetes while adding a control plane designed for high-density agent workloads.
Standard Kubernetes is described as being optimized for thousands of long-running services. Agent Substrate, by contrast, is designed for the traffic generated by millions of sub-second tool calls that could otherwise overwhelm a standard control plane. Its control plane is designed to handle hundreds of millions of registered agents.
Agent Substrate retains Kubernetes-based scheduling and horizontal scaling with declarative configuration. Working with Agent Executor, Google says it can provide a foundation for large agent deployments while preserving an organization’s place within the Kubernetes ecosystem.
Available in Preview
Google is releasing Agent Executor in the open so developers can test its design with their own workloads and provide feedback. The project is available in preview, and Google says its concepts and implementation will continue to develop as users experiment with the runtime.