THE SHORT ANSWER
Yes, and more so than the model swap. Jakob Nielsen's expectation economy piece ends on the guideline to treat model changes as redesigns, with the August 2025 case as evidence: GPT-4o replaced by GPT-5, users revolted, the old model back within three days. His ledger is right: familiarity capital is banked by consistency and spent by redesigns, each surprise is a tax, and redesigns get a two-to-four-week A/B test and a classic path. The layer is wrong. The user of an agent product never sees the model; what they feel is the harness, the prompt, the skills, the retrieval budget, the tools. Lesson 08 of Becoming Forward Deployed Engineer has a prompt shortened on a Thursday, four lines removed, seven fewer answers out of 42, noticed two weeks later. Vivek Trivedy moved one model about thirteen points by changing only the harness while the two models in Lesson 06's review were three points apart. My own harness library has 51 commits since March 30, 30 with no name, none tested for two to four weeks, none with a classic path. By Nielsen's ledger each was a redesign. Put six lines under every harness commit: score, rows, name, undo, the previous version selectable for the test window, and a slice before the fleet. The model swap gets the same six; it is just the smallest redesign you ship.