Harness

The layer around the model, not the model itself: agents, evaluation, tool use, memory and context, behavioural control, human–AI interaction, and incentives.

AI Infra is how models are trained and served. Harness is how they are steered, measured, and made to act.

The first project here is LLM Advertising — advertising inside and around LLM interactions.

Later shelves (not started):

• Agents and tool loops
• Evaluation
• Memory and context
• Behavioural control
• Human–AI interaction