Harness engineering is the next frontier of performance optimisation in AI, and most of you are already doing it without realising.
What a harness actually is
Everything else in play other than the model — your CLAUDE.md, system prompts, skills, slash commands, memory layers, context window management, and so on. The list goes on.
If the model is the engine, the harness is the entire vehicle that turns raw output into useful capability.
People are building harnesses every day without naming the discipline. Naming it helps, because you’re more likely to start designing it like an engineer.
Interestingly, harnesses may have an equal or greater influence on performance than the model itself. The same model can swing 20–40 points on a benchmark depending on its harness. Smaller, well-harnessed models routinely beat larger, badly-harnessed ones. Resources spent on harness development currently outperform resources spent on parameters and model improvement, and likely will for a while.
Your harness is not a static artifact, and the model underneath will be swapped. Tool descriptions will be rewritten by the next model in the loop. Your judge will be a different model in six months. The context window will quadruple.
Harness engineering is not designing for the current frontier model, but rather for a moving target where every component — including the harness itself — is on an optimisation gradient.
It is a rich topic, because it encourages optimisation with existing resources and cultivates creativity and technical solutions, rather than offloading performance to the “next great model coming in 6 months”.
Harnesses are not only likely to change, but are expected to, so major AI labs have shifted development towards a meta-harness design. Meta-harness design forces different questions:
- Will this scaffold still make sense when the underlying model is 3x more capable?
- Which parts should be hand-written, which learned, which auto-tuned?
- What’s my interface, so I can swap models without rewriting?
- How do I version a harness whose pieces are themselves being co-evolved with the model?
The next frontier isn’t bigger models — it’s the ability to optimise with existing tools or systems. The engineers who design the systems around the models, and design those systems to survive the next model, will outperform the rest.
What harness engineering have you done that has optimised performance?