Agent Executor, Google’s distributed Agent Runtime

Google is open-sourcing Agent Executor, a distributed runtime standard for running AI agents reliably in production. It addresses key challenges of long-running agent workflows including durable execution with resume-after-failure, secure sandbox isolation, session consistency via single-writer architecture, connection recovery for disconnected clients, and trajectory branching via checkpoints. Agent Executor is harness-agnostic and supports LangChain/LangGraph, ADK, and Agent2Agent Protocol. It pairs with a new project called Agent Substrate, a Kubernetes extension designed to handle hundreds of millions of agents with sub-second tool calls, bypassing standard Kubernetes control plane limitations. Both projects are available now in preview on GitHub.