What is an agent harness?
An agent harness is the surrounding system that runs a model-driven workflow: it handles tools, state, execution limits, feedback, and recovery. The model is only one part of the agent. Harness design affects which tasks can be attempted and how success or failure becomes visible.
What does the source say?
Anthropic’s harness-design work examines long-running application development and evaluation. It provides a concrete example of changing the execution system around a model instead of relying on prompt wording alone.
How can you apply this?
Sketch one loop with input, decision, tool call, result, check, and stopping condition. Then simulate a tool timeout, an interrupted run, and an incorrect output. Specify what state survives each event and what the next run must know. Make the stop button and recovery owner as concrete as the happy path. Ask a teammate to explain the recovery path using only your sketch; revise any step that depends on hidden knowledge.
What should you watch out for?
A loop that keeps trying is not necessarily making progress. Bound retries, total cost, and the actions allowed after repeated failure.