What is Agent harness?

There is no formal specification for an agent harness, and the term is used loosely. The cited page is Anthropic's engineering write-up on long-running agents, and it does not give a strict definition either. It uses the word without defining it, calling the Claude Agent SDK a general-purpose agent harness and describing a harness design for tasks that span many context windows.

The page describes problems a harness has to handle once a task outgrows one context window. A new session starts with no memory of earlier work, agents can try to do too much at once, and they can declare a project finished early or mark a feature done without testing it end to end. Context management, such as compaction, helps the agent keep going without filling its window.

Anthropic's described fix is a harness design in which an initializer agent prepares a feature list, a script to start the app, a progress file and a first git commit. Later coding agents then take one feature per session, test it, and leave notes and commits behind. This is one vendor's example, not a standard pattern.