An open agent harness exposes its controls

In a Peter Yang interview, Nous Research co-founder Karan Malhotra argues that an open agent harness changes how a model behaves by combining prompts, persistent memory, skills, tools, and user context around the model rather than relying on its native chat or coding interface. Why it matters: The model may be the same, but the harness decides what it remembers, which tools it can use, what gets verified, and whose preferences shape the run. That makes the harness an operational trust boundary, not just a prettier interface.

Try this: Run one bounded task through a native interface and an alternative harness with the same model. Compare the saved context, tool permissions, review pauses, outputs, and data-retention defaults before treating either result as a model capability claim.