What is an agent harness?
An agent harness is the execution scaffolding that connects tool calls, memory, and a repeated execution loop to a language model (LLM), turning a model that only answers into an agent that actually performs work.
Even with the same model, the harness you wrap around it can change the agent's real performance and stability a great deal.
How is an agent harness structured?
It generally consists of a prompt that gives the model the situation, tool calls that invoke outside functions, memory that retains earlier context, and a repeated loop of "decide, act, check the result."
In multi-agent setups where several agents collaborate, an additional layer may coordinate them.
How is a harness different from an agent or a model?
A simple way to think about it: the model is the brain that makes decisions, the harness is the skeleton that turns those decisions into actions, and the agent is the finished system that combines the two to accomplish a goal.
With the same model, a weak harness means the tools aren't used well, while a well-designed harness makes up for much of the model's weaknesses.
What is harness engineering?
Harness engineering is the work of designing and improving which tools to expose and how, how to manage context, and how to retry after failures, in order to raise an agent's performance and stability.
Recently, it has been emphasized that harness design shapes the quality of real results as much as making the model bigger.