Introducing Agent Executor, Google's Distributed Agent Runtime

As models and harnesses advance, agents now perform complex tasks that run for hours or even days. However, this has exposed operational challenges: long-running agent workflows are fragile and extremely difficult to manage reliably and efficiently in production environments.

Agent Executor is an open-source agent execution runtime standard released by Google, designed to solve these problems by providing the following features:

• Durable execution: Automatically provides resilience through event logs and snapshots, allowing resumption after interruptions such as failures or HITL confirmations
• Secure isolation: Isolates components in secure sandboxes when agents generate code or handle multiple tenants
• Session consistency: Maintains shared state consistency in distributed workflows through a single-writer architecture
• Connection recovery: Allows clients to reconnect to agents after network failures and restore responses from the last sequence
• Trajectory branching: Enables testing different decision paths without losing context, via checkpoints

Federation of Deployment Models

Agent Executor connects various deployment models, allowing enterprises to mix and match according to their needs. It supports Google's latest agents such as Antigravity 2.0 and Deep Research, custom managed agents via the Managed Agents API, and self-built agents made with LangChain/LangGraph or ADK.

Enterprises can deploy agents on their own infrastructure to avoid vendor lock-in and maintain data sovereignty. Thanks to its harness-agnostic design, agents built with any framework can be used, and the entire agentic stack—including MCP, skills, and other agents—can run on one's own data plane.

Scalability on Kubernetes

Google also announced a new open-source project called Agent Substrate alongside this. It is a new abstraction layer built on top of Kubernetes that allocates/deallocates agents to compute capacity in real time, providing lower latency and higher efficiency. While standard Kubernetes is optimized to handle thousands of long-running services, Agent Substrate is designed to handle millions of sub-second tool calls.

Agent Executor is currently available in preview and can be accessed at the GitHub repository.