An agent runtime is the environment in which an AI agent, a program that takes a goal and autonomously uses tools to carry out tasks, actually executes. It provides the execution infrastructure for model calls, tool execution, state and memory management, permissions and access control, and sandboxed isolation, sitting between agent frameworks at the application layer and the underlying infrastructure.
Just as a conventional runtime supplies what a program needs to run, an agent runtime is an execution environment designed around the actions agents perform. Cloud providers and AI companies offer managed services and open-source runtimes under this name for deploying and operating agents, and the runtime also serves as the point where agent behavior is observed and policies are enforced.