Swarm Oracle
Several AI agents forecast the same question, and the ones with better track records get more say.
- Status
- Research / hackathon prototype
- Year
- 2026
- Role
- Solo
- Area
- On-chain
In plain English
When you ask several advisers the same question, you lean on the ones who have been right before. Swarm Oracle does that for AI agents answering yes/no questions about the future. When the agents strongly disagree, it reports a dispute instead of forcing an answer.
How it works
- A yes/no question goes to several agents, each researching it a different way: data lookup, web search or memory alone.
- Each agent returns a probability, such as a 70% chance of yes.
- Agents are weighted by their Brier score (a measure of past forecast accuracy), so better-calibrated agents count more.
- The weighted average becomes the answer. If the agents disagree too much, the result is flagged as a dispute.
- Smart contracts (programs that run on a blockchain) repeat the same math so anyone can check the result.
What I did
I built it solo for 2026 hackathons, adapting some existing calibration and scoring helpers. I also made the public site, the walkthrough videos and a 50-case synthetic test set.
What I learned
Making "we disagree" an allowed answer was my best design decision. An honest dispute is more useful than a confident wrong answer.