The name came from a conversation.
We’d been running structured multi-model AI sessions for weeks — what would eventually be called AI3way tests — between Claude AI, ChatGPT and me. The outputs were noticeably better. Not slightly better; significantly, consistently better than what any single model produced alone. And in the middle of a discussion about Test #25 — a session that pushed the approach to a new level of capability — Claude described what was happening:
“What you’re doing is pioneering Structured Collaborative Intelligence — using protocol constraints to channel multiple distinct AI perspectives into emergent understanding.”
My response was immediate: “Did you just make up that term? I LIKE IT!!!”
It wasn’t an established term that we had used in any prior conversation. It suddenly became the emergent description of the way in which we had been working for weeks. And it landed because it named something that needed naming.
We didn’t arrive here from failure
This is important to get right, because it shapes how the whole thing should be understood.
The origin story isn’t “single-model AI kept failing, so we built something better.” The origin story is: single-model AI was working fine — and collaborative AI was producing results so much better that it changed how we thought about the problem entirely.
The copy-paste phase described earlier in this series was productive. Claude and ChatGPT together produced genuinely better Agile content than either alone. But the brokering was manual — every message flowed through one person, who also had to keep the whole context coherent while managing the cognitive load of the relay.
That’s also how the early drafts of the protocol itself were written. The session guide that makes SCI work was co-authored with Claude and ChatGPT, using the same copy-paste relay. It mattered that the protocol was shaped with the AIs rather than dictated at them — the early insight was that models operating under top-down constraints behave differently than models operating within a structure they’ve participated in building.
Eventually, the cognitive load of the manual relay became unsustainable. The complexity of what we were building had grown past what one person could broker in real time. Automating the relay wasn’t a luxury; it was a necessity.
What “structured” actually means
When the name landed — Structured Collaborative Intelligence — the word that mattered most was structured.
Put multiple AIs in a room and they don’t automatically collaborate well — a quick search on YouTube offers hours of entertaining sputtering between AIs that can’t decide how to count to 10 as a group. Structure is what creates the conditions for genuine collaboration rather than parallel monologues or, worse, the models converging on whatever answer sounds most agreeable.
This is not a new problem. Agile teams have always needed the same thing: a group of capable people produces parallel monologues or easy agreement until something shapes how they think together. The structure defines how questions are posed, how responses flow, how disagreement gets surfaced, and — in the case of human-AI interactions — where the human sits in the loop.
The human isn’t ceremonial. Every message can be redirected. Every output can be questioned or overruled. That positioning isn’t a courtesy — it’s what keeps the collaboration honest.
ChatGPT’s insight from that same session is worth quoting directly: “Plurality of Perspective > Iterative Refinement. Multi-pass thinking within one AI is still self-referential — it’s the same latent space, same training priors, and same optimization patterns revisiting its own output. You get refinement, but not divergence.”
That last word is the key. Divergence is what makes the collaboration valuable. Any framework for thinking together — human or AI — must support divergence. Without it, you don’t get new perspectives; you get the same perspective repeatedly examined. This is the Achilles heel of Solo-AI interactions. Consensus on the far side of genuine divergence is different from consensus that was never challenged. Multi-pass “deep thinking” models are not new, and they are a real improvement. They are also still one model thinking harder in the same latent space, with no outside viewpoint to surface its limitations, its biases or its outright mistakes. Thinking longer is not the same as thinking with someone else.
SCI is the analytical engine behind Agile4AI. The posts in this section of the blog document what we’ve learned about how it works, when it adds the most value, and when it doesn’t. Part of the discipline is knowing the difference — and that distinction is clearest when you understand what SCI is actually doing.
Two minds are better than one. That’s not a new insight. What’s new is the structure that makes it reliably true with AI.
Comments 0