Google has introduced Agent Executor, an open-source runtime standard for AI agent execution, resumption, and distributed deployment.
The project is aimed at long-running agent workflows that can continue for hours or days. Google said long-running execution requires the ability to resume after outages or agentic interruptions, including human-in-the-loop confirmations. It also cited client disconnections and distributed execution as challenges for long-running workflows.
A runtime for long-running agents
Agent Executor includes durable execution features for agents, agent harnesses, tools, skills, and sandboxes. It uses an event log and snapshotting to allow workflows to resume after outages or interruptions, including human-in-the-loop confirmations.
LangChain’s 2026 State of Agent Engineering report found that 57.3% of surveyed respondents had agents running in production, while 30.4% were actively developing agents with deployment plans.
The runtime also supports connection recovery. When a client disconnects during a long-running workflow, Agent Executor allows the client to reconnect and receive responses from the last sequence it had seen.
Another feature, trajectory branching, allows checkpoints to be used when testing different paths in an agent workflow. This lets agents branch from a previous checkpoint while preserving context and state.
Security and state management
The runtime includes secure isolation through sandboxed components. Google said this is intended to limit harmful side effects and reduce the risk of malicious activity affecting the wider service. It cited code generation and multi-tenant user data handling as cases where sandboxing is useful.
Google’s GKE documentation describes Agent Sandbox as being built for AI agent runtimes where untrusted, LLM-generated code must run in an isolated environment. It supports kernel-level isolation through GKE Sandbox and can also work with Kata Containers.
The same documentation says GKE Agent Sandbox uses a default-deny network security posture for sandboxed environments. The policy blocks untrusted code inside a sandbox from accessing unauthorised internal networks or the GKE control plane by default.
Agent Executor uses a single-writer architecture to manage shared session state. The architecture limits conflicting updates when several components in a distributed workflow attempt to modify the same session at once.
Deployment across agent frameworks
Agent Executor can connect with several deployment models, including Google Antigravity, Google-built agents such as Deep Research, and custom agents managed through the Managed Agents in Gemini API. It also supports agents built with LangChain, LangGraph, and Google’s Agent Development Kit.
Support for the Agent2Agent Protocol allows it to work across different agent frameworks and deployment environments.
Agent Executor can be deployed on self-managed infrastructure. Google said this allows proprietary workflows to run within an organisation’s own compute environments and sandboxes.
The runtime is also agent-harness agnostic. Developers can use their own harnesses or agents from other vendors. They can also run model context protocols, skills, and other agents within their own data plane.
A Kubernetes layer for agent workloads
Google also announced Agent Substrate, a related open-source project developed with the Google Kubernetes Engine team. The project adds an agent-focused layer on top of Kubernetes for running large numbers of agent workloads.
Google said agent workloads differ from standard cloud services because they can involve short bursts of activity followed by longer idle periods. GKE Agent Sandbox integrates with Pod Snapshots so idle agent workloads can be suspended and resumed in seconds.
Standard Kubernetes is designed for thousands of long-running services, according to Google. Agent workloads can involve millions of short tool calls. Agent Substrate moves agents onto and off ready compute capacity in real time, using a smaller control plane for this execution pattern.
Sandbox allocation
GKE Agent Sandbox’s warm pool can allocate 300 sandboxes per second per cluster, according to Google. The company said 90% of allocations are completed in 200 milliseconds. Google said the warm pool is intended to reduce cold-start latency when new sandbox instances are needed.
Agent Substrate combines secure runtime and snapshotting features from sandbox infrastructure. It also uses Kubernetes-based scheduling and horizontal scaling. Google said the project introduces a control plane designed to handle hundreds of millions of registered agents while remaining within the Kubernetes ecosystem.
See also: Google adds Android app generation and Managed Agents to Gemini developer tools
Want to dive deeper into the tools and frameworks shaping modern development? Check out the AI & Big Data Expo, taking place in Amsterdam, California, and London. Explore cutting-edge sessions on machine learning, data pipelines, and next-gen AI applications. The event is part of TechEx and co-located with other leading technology events. Click here for more information.
DeveloperTech News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.