The strongest version
Intelligence is an assembled system, not a property of any single network. Memory, planning, tool use, reflection, and reasoning composed around a model — with the model as one component among several — is the route to general capability. The lesson of agent frameworks is that the same base model becomes dramatically more capable when embedded in the right architecture of loops and stores.
Who holds it
The lineage runs from the classical cognitive architectures (SOAR, ACT-R) through modern agent scaffolding. The contemporary version is less about a fixed blueprint and more about the claim that orchestration is where the remaining capability lives.
The load-bearing assumption
That the composition is doing real cognitive work — not merely papering over a weak base model — and that the assembled system exhibits capabilities none of its parts has alone.
Falsifiers
If end-to-end trained models absorb the functions of the scaffold and outperform assembled systems (memory, planning, tool use becoming native rather than orchestrated), the architecture-as-system thesis weakens. Health is holding: scaffolding clearly helps today, but the trend line for “native” capability is the thing to watch.