If you’ve read the latest revision to Ethan Mollick’s guide to which AI to use, it starts with a striking example of how output from the same (frontier) model can vary depending on what harness you use to run it. His point is that the choice of harness matters just as much as the choice of underlying model.
Mollick also gave us the concept of “jagged frontier” (1) - the idea that AI is brilliant at some tasks and embarrassingly bad at others, and that the boundary between those two zones is unpredictable. You can’t just say “AI is good at writing” or “AI is bad at math.”
The harness argument now points at a new frontier - the “harness frontier”. When people say “I tried AI and it couldn’t do X,” the failure is often the harness. The same model hits a completely different capability boundary depending on what harness it’s in. Claude Opus in a chat window vs. Claude Opus in Claude Code vs. Claude Opus in Cowork are categorical, not incremental, differences.
The harness frontier is more predictable (because it’s baked in by product design) but not less jagged. You still have to discover it and then decide how to compensate for it via context design (prompt, skills, MCPs, etc.). Frontier labs marketing often obscures the harness frontier. Every harness describes its purpose in broad strokes (code, work), and you’re supposed to discover yourself the actual limits of ‘ask anything’.
Discoverability is an existing software product design problem. But in the AI products it’s magnified by the cumulative effect of these two overlapping frontiers in a non-deterministic space. Not knowing what you don’t know leads to optimization problems. AI will satisfice (2) most people, while obscuring its potential. To make things even harder, the harness frontier is a moving target, just like the frontier for the underlying models. You can’t build your mental map of the frontier once and call it a day.
So what can we do? While things are still moving at lightspeed, continue investing time in rebuilding tooling around your preferred model+harness. Try out new harnesses as they come out and don’t get attached to your current setup. Just don’t fall into the bikeshedding (3) trap.
PS. Read Ethan’s article (4), especially if you’re just moving past simple prompts. It’s the best map of the current landscape.
References:
(1) The Shape of AI: Jaggedness and Bottlenecks
(2) Satisficing