AI coding tools are making the model too easy to overvalue.
The model matters, obviously. But the daily experience around it decides whether the tool becomes part of engineering work or another impressive tab people stop opening. Same model, different shell, very different result.
Cursor's agent-first workspace changes visibility. You steer more work in parallel, inspect rendered behavior, and spend less time pretending that every valuable interaction begins with reading a file line by line. That can be powerful. It can also make experienced engineers uneasy because control feels different when the IDE becomes an orchestration surface.
Terminal-native tools create the opposite kind of friction. They can be powerful and alienating at once. Less "I am editing code" and more "I am supervising invisible labor." Some people will love that. Others will feel the abstraction eating the part of development where they build trust.
Then the product economics show up. Rate limits, token burn, prompt caching, and latency are not side details. They are part of the tool. A brilliant model that hits a wall 10 minutes into paid work is not a workflow. It is a preview with invoices.
The durable question is fit for task. Which tool keeps enough of the work visible, fast, and reviewable for the risk in front of you.
The smartest model still has to live somewhere.
Related episode: AI Coding Is More About The Experience Than The Model.
