We started with a YouTube channel and a stubborn conviction.
In 2024, Agile had a bad reputation — and not entirely without reason. Twenty years of misapplication, certification mills, SAFe diagrams that occupied entire walls, consultants who turned a mindset into a compliance exercise and a cash cow. The disenchantment was real, and it was earned.
Underneath all of that, the underlying insight — that complex work can’t be managed the same way as simple work, that the people closest to the problem need to be trusted to solve it, and that it’s not possible to manage unknowns with a Gantt chart — that still held. It hadn’t become untrue just because it had been misused. And the misinformation had hidden the real value of Agile at a great cost to companies and workers alike.
So our goal was to set the record straight. To separate the Agile mindset from the Agile industry. And to answer one question: does any of this apply to AI?
A daily commitment and a workflow problem
We launched the YouTube channel on September 10, 2024. One video a day. That’s a massive content commitment, and it required a workflow — so AI became part of our process immediately.
Our approach: start with a brain dump. Ideas, structure, goals for the video — all of that came first, authored entirely by hand. Then we handed it to Claude for a first draft.
What Claude did well was striking. It could take a stream-of-consciousness idea dump and produce something structured and coherent that preserved the original intent, tone, and voice without flattening it. Not just following the input literally — interpreting it, enriching where it made sense, holding back where it didn’t. The first draft felt like the original idea, only organized.
And then there was a gap. The drafts were good — and a little heavy. Accessible to people already in the world, but not quite reaching the broader audience we needed to reach.
That’s where ChatGPT came in.
Two halves of a whole
Handing Claude’s draft to ChatGPT revealed something we didn’t expect. ChatGPT brought a different set of strengths: it distilled, tightened, and made things crisper from a different point of view. It added factual texture — specific details that grounded the ideas in concrete reality. The output was easier to follow without losing the substance underneath.
The combination was clearly better than either alone. Claude shaped the idea faithfully; ChatGPT polished the delivery. And then, bouncing the result back through Claude for a final pass often surfaced one or two sharp observations that improved the piece further.
Three voices. A richer result.
The question
The workflow worked. And it was exhausting.
Every message flowed through us — writing a prompt, reading the response, copying it over, writing the next prompt, reading the next response, copying again. Back and forth, back and forth. The manual relay between two models that clearly complemented each other was productive and relentless.
At some point, we asked the obvious question: what if they could just talk to each other?
Not a grand theory. Not a product vision. Pure pragmatism — frankly, laziness. The cut-and-paste was a lot of work, and the obvious improvement was to let AIs collaborate until they had something they could agree was of good value to present back for human evaluation.
But the question turned out to be bigger than it looked.
Because once we asked how to get two AIs to collaborate directly, we weren’t asking a technical question anymore. We were asking what collaboration even means between AI participants. Whether structure matters — without which amusing YouTube videos get produced of AIs engaging in never-ending back-and-forths. And, what happens to the output when two models with different strengths engage the same problem without a human translating between them?
Those questions are what the rest of this blog is all about.
Comments 0