In the context of AI agents, a harness is the software infrastructure that surrounds a language model and allows it to act as a functional agent. It is not about the model itself, but rather the set of components that manage the workflow: the construction of prompts, access to external tools, memory, control of execution steps, and the capture of results.
Its importance lies in the fact that an agent's performance depends as much on the underlying model as on the harness that orchestrates it. The same model can offer very different results depending on how context is provided or how its tool calls are managed. Therefore, when comparing agents in benchmarks such as SWE-bench, it is essential to specify which harness was used.
A typical harness is responsible for:
- Iterating on the agent's actions until the goal is reached.
- Connecting the model with tools (terminal, browser, APIs).
- Logging traces to evaluate and debug behavior.