Six models, one variable, the same result

What stopped me reading 'The Harness Effect' wasn't the 41% figure — it was that they tested six different models, from Claude Sonnet 4.6 to Palmyra X6, and changed only the orchestration layer. If the gain shows up even when the model is held constant, the gain was never in the model.[1]

The numbers hold up: cost per task fell from $0.21 to $0.12, tokens dropped 38%, and quality actually rose from 0.78 to 0.81 — it got cheaper and more accurate at once. Seeing a 33-61% gain across all six models tells me this isn't one model's special talent; it's the outcome of an architectural decision.[1]

MCP dropping session state is the infrastructure half of the same story

The second story I read the same day was MCP moving to a stateless server-side session ID. Arcade founder Nate Barbettini's description explains exactly why the orchestration layer has been straining: 'Every one of those machines has to know about a session ID that some other machine handed out. It's not impossible, but it's a serious pain.'[2]

On one side, an orchestration pattern that cuts spend; on the other, a protocol update that makes running it across distributed infrastructure easier — both are investments in the same layer. The developer's job is no longer picking the model; it's building this layer correctly.[1], [2]

What I said on July 16 got measured today

In my July 16 column I said the model layer is commoditizing and value is moving up.[3]

On July 19 I wrote that the agent layer had genuinely gotten cheaper.[4]

On July 20 I wrote that Current AI confirmed it on the nonprofit side.[5]

Today's Writer-authored vendor research supports the same claim with a controlled experiment across six models; it is a measurable vendor finding, not independent validation.[1]

I'll confess: this didn't surprise me, I'm just glad to see it confirmed. My next test is how fast open-source communities copy these harness patterns — if the gain is really architectural, we should see it show up in three separate open-source projects within a month.[1]