It was late, and for the first time I can remember, the fan on my MacBook turned on. 128 gigabytes of RAM, and I had never once heard it strain. Last night it strained. I had a system running that was building a presentation, and across the run it had spun up 178 AI agents, while my laptop drove a live dev server and a headless browser screenshotting every slide, over and over. The machine finally had something to complain about.
Here is what it was doing.
The System
For weeks I have wanted one thing: to build a full slide deck in a single pass and have it come out at the quality bar I would accept by hand. I am a former designer. I have a high standard for this, and AI deck tools have never met it. The decks come out generic. So instead of asking a model to "make me a presentation," we built a process around the model.
The deck is code. Every slide is a React component in our own design system. The system reads a plan, fans out one agent per slide to write the component, assembles them, then renders each slide in a browser, screenshots it, and grades the screenshot against a written rubric and a set of hand-picked reference slides. Where a slide misses, another agent rewrites it, and the loop runs again. Render, screenshot, grade, fix. The same loop a designer runs in their head, turned into something a fleet of agents can run in parallel.
Forty-Five Minutes
Last night we proved it. In a single run of about forty-five minutes, the system produced a deck of thirty-six slides that was most of the way to finished. Not a rough draft. A deck I could open, read end to end, and hand-correct slide by slide instead of rebuilding. Thirteen percent of the slides needed a human pass. The rest were good enough to keep.
It cost just under 112 dollars in model usage, about three dollars a slide. That is real money, and I want to be honest about it. This is not free, and it is not magic. It is a machine that does a lot of careful work quickly, and you pay for the work. It also took more than one attempt to get there. Earlier runs stalled and broke in ways we had to fix before the clean one went through.
The honest part is that it does not get to 100. A person with taste still finishes the deck. The system gets you most of the way, fast, and then the human closes the gap. That division of labor is the whole point.
The Jagged Frontier
Ethan Mollick, the Wharton professor, has a phrase for where we were working last night: the jagged frontier of AI. AI is strangely good at some things that look hard and strangely bad at some things that look easy, and the line between them is jagged, not smooth. The phrase did not come from a lab. It came from a field experiment on management consultants, measuring where AI helped their work and where it quietly hurt it.
That is exactly the terrain we were on. The system can write thirty-six slide components in parallel without breaking a sweat, then completely misjudge whether a card looks balanced. So we built guardrails around the parts it gets wrong and let it run free on the parts it gets right. Finding that line, for a specific kind of work, is the actual job.
Days With Clients, Nights on the Edge
This is what we do at Gadoci. During the day, we sit with clients and turn AI tools into real operational work. That is the business. The reason we can do it is that someone is out on the edge at night, finding out what these tools can actually do before anyone is paying for the answer.
That someone, most nights, is me. Last night it was a deck-building system and a laptop fan I had never heard before. Some nights the experiment fails. Last night it worked, and it worked well enough that I am writing about it.
The frontier is jagged, and it moves. The only way to know where it is this week is to go out and touch it. We will keep doing that, and we will keep bringing back what we find.