Agent harness runs an agentic loop

Earendil describes an agent harness as the software environment around a model: system instructions, available tools, an agentic loop, and a layer that can connect the workflow to different models. Why it matters: The model is only one part of an agent product. The harness decides what context enters the task, what actions are available, when the model can retry, and where a person can inspect or stop work.

Try this: Inspect an agent's system prompt, tool permissions, loop stop condition, and model-routing policy before you judge its behavior.