AI engineering paradigms: Anthropic harness and Loop/Graph Engineering debate

Anthropic reports the model 'harness' layer is shrinking; new layer behaves like a strategy coach rather than a fixed pipeline.
Key insights
Anthropic says the outer 'harness' layer is thinning as model capabilities grow
An internal Anthropic discussion reported that the 'harness'—the layer of code wrapped around models that encodes assumptions about what models cannot do—is being reduced. The post states that many of the harness's assumptions are expiring as model capabilities advance. It contrasts the old harness, described as a fixed assembly-line pipeline with predetermined steps, with an emergent layer that acts more like a coach, selecting strategies and deciding how to approach tasks rather than enforcing a fixed sequence.
Community pushback questions proliferation of new engineering labels like Graph Engineering
A separate post reacted to new engineering nomenclature, questioning whether Loop Engineering is already outdated and asking what 'Graph Engineering' is. The reaction called for restraint in inventing new AI engineering concepts, criticizing them as token-consuming and of limited utility in practice.