← All Posts

March Madness

How 16 AI Agents Predicted Every March Madness Game

March 17, 2026

The odds of a perfect March Madness bracket are roughly 1 in 9.2 quintillion. No human has ever done it. But what if you replaced gut instinct with 16 AI agents simultaneously analyzing KenPom ratings, betting markets, injury reports, tempo stats, and real-time X/Twitter sentiment? That's exactly what we built.

We used Probe AI's Deep Research to generate a complete 63-game bracket for the 2026 NCAA Tournament. Five queries, $10 in credits, and roughly 8 minutes of total research time. The result: a fully structured bracket with confidence scores, upset flags, and reasoning for every single pick. You can track the bracket live here.

The Methodology

We ran 5 Deep Research queries ($2 each)—one for each of the four regions (South, East, Midwest, West) and one for the Final Four through the Championship. Each query deployed up to 16 parallel agents that simultaneously searched the web and X/Twitter for the latest data.

The prompts asked for structured picks with confidence percentages, upset flags, and reasoning. We didn't cherry-pick or override any results. What the AI returned is what we published. The raw picks were extracted programmatically and stored as structured data for the live bracket tracker.

Data Sources

KenPom Ratings: Adjusted offensive/defensive efficiency, tempo, strength of schedule. The gold standard for college basketball analytics.
Betting Markets: Vegas lines, Polymarket odds, and public money percentages. Markets aggregate information from thousands of bettors and sharps.
Injury Reports: Real-time injury status from official team sources and beat reporters. A single injury can swing a game by 5+ points.
X/Twitter Sentiment: What insiders, analysts, and the betting community are saying in real-time. Early indicators of line movement and narrative shifts.
Historical Matchups: Seed-vs-seed historical performance, conference strength, and tournament experience for coaches and rosters.

Bold Predictions

The AI didn't play it safe. It flagged several upset picks where the data suggested the lower seed had a genuine edge. Here are the boldest calls:

VCU over UNC: KenPom defensive efficiency edge, UNC's turnover rate exploitable by VCU's press, and declining CT confidence in UNC after late-season losses.
High Point over Wisconsin: Conference tournament momentum, Wisconsin's 3-point shooting regression, and sharp money moving toward High Point on multiple books.
Utah State over Villanova: Tempo mismatch favoring Utah State, Villanova's road performance decline, and X/Twitter sentiment flagging a key Villanova rotation player as questionable.

Why Upsets Get Flagged

Every pick includes a confidence score from 0–100%. Picks below 65% are automatically flagged as potential upsets. But the interesting part is whythe AI assigns low confidence to higher seeds. It's not random—it's a convergence of signals.

When KenPom says a matchup is close, Vegas lines are tight, and X/Twitter is buzzing about a team's vulnerability, the AI recognizes that pattern. It's doing what a good human analyst does—triangulating across data sources—but at a scale and speed that's impossible manually.

Track the Bracket Live

We built a live bracket tracker that grades every pick against actual results. It updates automatically—you can see which picks hit, which busted, and whether the perfect bracket dream is still alive. The page also shows confidence scores, upset flags, and reasoning for every game.

We're also posting updates on @tryprobeio on X throughout the tournament—tracking record, highlighting upsets, and celebrating (or mourning) each result.

More from the blog

Want to run your own deep research? New accounts get $5 in free credits—enough for one Deep Research+ or 10 Quick Research queries.