In December I took part in Agentathon 2025 at Malla Reddy University in Hyderabad. More than 2,300 developers built at the same time, and the event set a Guinness World Record for the largest agentic AI hackathon. The record belonged to the room. What I took home was smaller.
I built fraud-detection agents for public-spending transparency: a multi-agent system that reasons over spending records and keeps every decision in a Neo4j knowledge graph. Each decision is stored as a versioned record. When a new finding contradicts an older one, the older record is flagged rather than overwritten. By the end the graph held more than 500 nodes.
The part worth keeping
The contradiction flag. Agents that summarise financial records sound confident by default. If a later agent reads a transaction differently, the easy design lets the newest answer win, and the disagreement disappears from the record. Keeping both versions, with a flag between them, turns the disagreement into something a person can check.
The same concern shows up in a proposal I co-wrote later in the year, AetherState: cached memory should keep a pointer back to the evidence it came from, so you can tell when it has gone stale.
The record was a bonus. The lesson was humility.