nikolas.sapa
November 18, 2026

52 agents of recon before a hackathon

A hackathon has two parts people treat as one. The build, which happens inside the window. And everything you can know before the window opens.

The rules for the Gen AI Summit Hackathon were clear: teams of 1 to 5, ideas and boilerplate allowed beforehand, a pre-built product gets you disqualified. You build only during the 36 hours.

So I did the part that was allowed. I ran 52 agents on the event before it started and collected the output into 56 notes. It is public: genai-summit-hackathon-recon.

What it covered

Four areas.

Judges. Dossiers on the people who would score us: what they had said publicly, what they cared about, what they had funded or built.

Sponsors. What each sponsor sells, what a team building on their stack would look like, and what their terms said.

Mechanics. The timeline, the submission requirements, the judging criteria, the finals format. The criteria were published as usefulness, craft and "aura points". Submissions closed Sunday at 18:00 with a repo, a demo video and project details. Semifinalists pitched for 3 minutes. Four finalists then pitched again on Monday at SNFCC.

Past winners. How earlier editions were won, down to the code where it was public.

The pattern it found

The most useful note was not about any one judge. It was the overlap.

Reading across the dossiers and sponsor material, the same architecture kept showing up: a policy-gated agent. The model fills typed fields, a deterministic core does the math, actions are logged, and anything that writes to the world needs a human to approve it. A previous Athens winner, Keel, summed it up as "dollars never come from the model."

Different judges, different words, same idea: checked, not trusted.

What it got right

Three predictions held.

The finals format. The real gate was the Sunday cut to four finalists. The Monday stage was the showcase. That is how it played out.

Judges weighting verifiability. FIREYE, the project Marco Serat and I built, ended up with that shape. Resident alerts are deterministic. The agents explain with scripted replies, with no live AI-generated evacuation orders. Drone tasks need human approval, enforced by the backend. That made it to the final four.

Lars Rasmussen. The recon had a dossier on him before the event. He was on the judging panel for the finals.

What it got wrong, or never tried

The recon ranked candidate verticals: a Greek voice agent, a maritime claims agent, a property-transfer copilot. We built a wildfire command console. None of the ranked ideas made it to the build. Ideas are cheap before the event. What the briefs and the room look like on Saturday decides more.

And recon cannot give you the wow moment.

It can tell you what judges reward. It can tell you the format and the clock. It cannot tell you which 30 seconds of your demo will make a room lean forward. That comes from the project and the pitch, and you only find it by building and rehearsing.

We finished 4th of 47. The post-mortem there says the ceiling was the project and the presentation, not execution volume. The recon did its job: it told us the format, the people and what they reward. It could not write the moment.

Would I do it again

Yes. It is allowed, and it meant I walked in knowing the format, the people and the criteria.

Next time I'd spend less of it on who the judges are and more on one question: what single thing could we show them in 30 seconds?