Spona

What it does

The track brief says the University of Novi Sad has thousands of researchers and NTP Novi Sad hundreds of companies, yet they rarely work together. Spona is a team of agents on Mozaik that reads the whole EINSTEIN dataset (1,047 researchers, 10,779 papers, 153 companies with 489 challenges, about 7M tokens) and gives every company challenge an outcome with its reason: a written proposal, a close match, no close match at the university, or not a research problem. For the best pairs it writes an industrial PhD topic or a one-page joint project, and every fact about a person or company cites the record it came from. In the recorded run (5 min 48 s, about $1.52 of Claude credit plus the Gemini free tier) it wrote 4 proposals for 4 companies, its critic sent back 13 evidence claims and 20 draft statements that went beyond their sources, and it found 64 research challenges with no close match in the university's own published work, listed by sector for EINSTEIN. Whether these pairs would actually meet isn't measured yet, since the data is synthetic; the README says how to measure it. The live page replays the run and lets you pick any company problem to see who could solve it and why, or why nobody does.

How the agents run concurrently

Twelve participants join one Mozaik runtime and react to events on its bus; nobody calls anybody in order. Three Industry Analysts each read a third of the companies at the same time. As each batch lands, the Critic checks the cards, the Matchmaker searches a BM25 library (no model) and re-ranks pairs, and three Research Profilers, each owning whole faculties, start evidence cards for the best pairs while the analysts are still reading. The Critic reviews every evidence card as it arrives and sends unsupported claims back to their author, who revises while everything else keeps going. The Writer drafts all chosen proposals at once and the Critic reviews each draft as it lands. Shared state is one board in Mozaik's RuntimeState: challenge cards, pairs and their fit, evidence cards and claims, proposals, and every critic intervention. The Coordinator splits the work, opens the evidence budget as analysis progresses so the first results can't use it all, and picks the pairs once the ranking settles. In the recorded run 84 agent loops ran, up to 19 at the same time. The Gantt chart in the README shows every loop, and each loop streamed live to Mozaik cloud.