Addy Osmani

"Agent = Model + Harness. If you're not the model, you're the harness." Addy Osmani credited that equation to Viv Trivedy in Agent Harness Engineering (19 April 2026), then reformulated it: a correct model in an excellent harness beats an excellent model in a bad harness. He is an engineer at «Google», and the same claim runs under most of what he publishes. The leverage sits outside the model.

The second half of the argument is the uncomfortable one. In Cognitive Surrender (5 May 2026) he separates cognitive offloading, delegating the how while keeping judgment, from cognitive surrender, "borrowing the model's confidence as substitute for personal understanding". What decides which one happens is the operator's posture, not the tool. He names comprehension debt as the widening gap between code volume and human understanding, and argues that throughput metrics cannot distinguish understood work from rubber-stamped work. His remedies are deliberately small: pull requests of roughly 100 lines maximum, regular unassisted solo keyboard time, five personal heuristics and six structural guardrails.

The line he drew in November 2025, that «vibe coding» is not AI-assisted engineering, keeps getting redrawn at higher altitude: conductor versus orchestrator, then «The New SDLC With Vibe Coding», then «Loop Engineering» in June 2026, then the constraint ring diagram of 30 July 2026 with its back-pressure rule, that a loop is granted only the autonomy you can verify cheaply and reliably. The fixed point across all of it is that the human engineer keeps responsibility for architecture.