Glossary
Agent Harness
The software layer wrapped around a model that turns raw text generation into a working agent: the tool-execution loop, permission and approval gating, context and compaction management, and the guardrails that decide what the model's output is allowed to touch. The model generates text; the harness decides what that text can touch.
Frontier model capability has become a weaker predictor of agent performance than harness design. The same model, run through different harnesses, has been reported scoring anywhere from the mid-40s to the low-80s percent on identical SWE-bench-style tasks — the spread comes entirely from scaffolding, not the model.
If two teams use the same model, why would one get much better agent results?
Because the harness, not the model, often decides the outcome. The tool-execution loop, permission gating, and context/compaction management wrapped around the model determine what it's allowed to do and how much of its context stays useful. Reports of the same model scoring anywhere from the mid-40s to low-80s percent on identical benchmark tasks, depending on harness alone, back this up.
Where this comes up
Context Engineering for Production AI Agents
Need more than a definition?
If this term describes a problem you have, find out where you stand with a free scored diagnostic.