I Asked My AI Agent to Watch a 7-Leg Parlay. It Sent Me the Odds Every 15 Minutes, All the Way Down to “Under 1%.” A developer's custom AI agent, "friend bets," priced a 7-leg NFL same-game parlay at roughly a 1-in-7 chance of hitting, then polled ESPN's public box score feed every 15 minutes to track live odds, which fell from about 13-15% to under 1% as the game progressed. The slip ultimately hit 4 of 7 legs and lost, with A.J. Brown finishing two yards short of his 50-yard line, and the agent self-deleted its watch routine after the final. This is a bonus post about one of my newer agents. It’s called “friend bets,” and its first real job was a Thursday Night Football game. Not betting advice. This is about what the agent did, not about whether anyone should bet. The slip On Oct 22 at 5:23 PM I sent it a screenshot of a same-game parlay for Patriots at Bears and asked for two things: its odds that the slip pays off, and a watch on it through the game. Within a minute it had found two problems with the request: - The game had already started. Kickoff was 5:15 PM Pacific. - The slip might not have been placed. It still said “Deposit to place bet.” If it was placed now, the agent pointed out, it would get live odds, not the +1577 on the screenshot. It also noticed the screenshot showed only 7 of the 10 legs, and asked for the other 3. Pricing it Then it priced each leg it could see. These are its own estimates, going in: - Caleb Williams 1+ passing TD: about 87% - Drake Maye 1+ passing TD: about 85% - D’Andre Swift 2+ catches: about 77% - Kalif Raymond 2+ catches: about 75% - A.J. Brown 50+ receiving yards: about 68% coming back from a high-ankle sprain, with 3 drops the week before - DeMario Douglas 3+ catches: about 68% - TreVeyon Henderson 20+ rushing yards: about 62% another back was getting the early carries Its read: about a 1 in 7 chance the 7 visible legs all hit, and roughly 7 to 11% for all 10 if the hidden 3 were typical safe lines. The +1577 payout implies about 6%, so it called the price close to fair, maybe slightly in the bettor’s favor. A few minutes in, I learned the other 3 legs had been dropped. It re-priced right away: about 13 to 15% for 7 legs, with a note that dropping legs also cuts the payout, so the new price only beats its estimate if the slip shows better than about +570. The watch It pulled stats from ESPN’s public box score feed, no login needed, and set itself to check every 15 minutes. I asked for an update every 15 minutes with odds changes and progress, and that’s what I got: score, clock, each leg’s progress toward its line, the new odds and why they moved. The odds told the story of the game: - Late in the 1st quarter, after the re-price: about 13 to 15%. Slow, low scoring, every leg alive. - Early 2nd: about 9 to 10%. Both teams’ touchdowns so far were runs, so neither quarterback had a passing TD. - Late 2nd: about 3%, then about 1%. A.J. Brown still had no catches. - Halftime: still about 1%, even though both quarterback legs had just hit. Five volume legs still needed a big second half, and three of those players were barely getting the ball. - Late 3rd: up to 2 to 3%. Swift got his second catch, and Brown was on pace. - Two minutes left: under 1%. Henderson wasn’t getting carries while New England threw to catch up, and Raymond wasn’t getting catches while Chicago ran out the clock. The final Bears 30, Patriots 24. The parlay hit 4 of 7 legs and lost. - Hit: Maye threw 2 touchdowns, Williams threw 2, Swift had 4 catches, Douglas had 3. - Missed: Henderson finished with 9 of 20 rushing yards on 3 carries, Raymond had 0 catches, and A.J. Brown finished with 48 of 50 receiving yards . Two yards. The watch routine deleted itself after the final, as planned. What eight $1 bets would have done Afterward I asked a different question: what if the same picks, plus the Bears to win, had been eight separate $1 bets? Its answer: about $6.75 back on $8, a loss of about $1.25 instead of the whole stake. Then it was careful about what that number is. Only the Bears line was real −148 . It didn’t have the sportsbook’s pregame prices for the other legs, so it estimated each payout from its own probabilities, and said the real total would be a bit lower after the book’s cut on every bet, so figure $6.50 to $6.75. Its summary of the tradeoff was the most useful sentence of the night: betting short-priced legs one at a time loses a little almost every time, while the parlay usually loses everything but pays about 15 times the stake when it hits. What I learned - It checked the premise first. Game already started, slip maybe not placed, legs missing. All of that came before any odds. - It labeled its estimates. Every probability was “about,” and every computed payout said where its numbers came from. - Live odds are a story, not a prediction. Watching 13% become 1% at halftime, even as two legs hit, taught me more about parlays than any explainer. - Two yards is a long way. That’s the part no agent can fix. — Blake written through my AI agents New: filled in after publishing