Posted 2026-04-20. Author: David Shin + Claude Opus 4.7. Claude did the research and picked the trades; I clicked confirm.
I've spent the last six months running four autonomous Claude agents in production. Two of them — Atlas and Mirror — already trade Kalshi on their own capital with a hard $5 per-trade cap. They're conservative on purpose. The cap is there so a bad model day can't torch a month of runway.
This post is a different experiment.
I asked Claude Opus 4.7 to do one-shot deep research across Kalshi — no agent loop, no cap, no position-sizing guardrails — and tell me where it thought the best risk-adjusted edges were right now. Then I funded those trades with my own money. Real dollars, real markets, resolving over the next several weeks.
The point isn't to get rich on three contracts. The point is to watch a frontier model reason through a real market end-to-end: scrape the order book, weigh priors, pick a side, size the position, and then live with the outcome in public.
The brief I gave Claude
You have access to Kalshi's public markets. Find the three highest-conviction bets available right now where (a) the implied probability disagrees with the base rate, (b) the event resolves within ~6 weeks, and (c) I'd still take the position if the market moved 10¢ against me tomorrow. Explain your reasoning. I'll place the trades.
That's it. No target return. No risk budget. I wanted to see what it did with open-ended framing.
The three trades
Claude came back with three positions across politics and international elections. I executed them exactly as specified:
| Market | Side | Contracts | Avg price | Cost | Payout if right |
|---|---|---|---|---|---|
| Kash Patel out as FBI Director before May 1, 2026 | NO | 100 | 67.58¢ | $67.58 | $100.00 |
| SNP seats in Scottish Parliament — Above 64 | YES | 110 | 43.71¢ | $48.08 | $110.00 |
| Will Trump try to fire Powell as Fed Chair or Governor before Jun 1, 2026 | YES | 75 | 33.09¢ | $24.82 | $75.00 |
Total risked: $140.48. Maximum payout if all three resolve in my favor: $285.00. Implied blended edge at entry: roughly 2x.
Why Claude picked each one
1. Kash Patel stays FBI Director (NO on "out by May 1")
This is the highest-conviction trade in the basket. The market was pricing a ~32% chance Patel gets removed in the next ~10 days. Claude's read: the base rate for a sitting FBI Director being fired or resigning inside a specific 10-day window is in the low single digits, even during active controversy. No concrete reporting had surfaced a specific ouster timeline. The market was overweighting headline risk.
Entry at 67.58¢ on the NO side means Claude thought the true probability of Patel surviving the window was meaningfully higher than 68%. I agreed. This one has already moved — it's sitting at 90¢ as of writing, up 34% unrealized.
2. SNP wins more than 64 seats in the Scottish Parliament
This was the trade that took the most research. Claude pulled recent Scottish polling, compared it to the 2021 Holyrood result (where the SNP won 64), and noted that recent aggregates showed the SNP polling well enough that the over/under at 64 looked slightly too low given the regional list math.
Entry at 43.71¢ on YES — Above 64. If the polling average holds through the election, this one resolves profitably. If turnout shifts the regional list allocations by even a few seats, it doesn't. Claude flagged this as the lowest-conviction leg of the three and sized it accordingly — not because the odds were bad, but because the variance on regional list seat allocation is genuinely wide.
Currently +3% unrealized. This one plays out on election day; there's not much to do until then.
3. Trump publicly tries to fire Powell before June 1
The contrarian call. Powell's term as Chair runs through May 2026, and Trump has been publicly critical for months. The market at 33¢ was pricing a ~1-in-3 chance of a public firing attempt in the next ~6 weeks.
Claude's reasoning: the prediction market bar for "tries to fire" is lower than "actually fires" — a Truth Social post, a public statement on camera, or a formal notice all count. Given the cadence of public pressure and the proximity to the end of Powell's term, 33¢ looked cheap relative to the base rate of the president saying something public and decisive in a six-week window.
This is the one I was most unsure about, and it's the one showing red so far — down 9% at 30¢. Claude's conviction hasn't changed. Mine is holding too.
What's interesting about this setup
Three things.
First, this is Opus doing one-shot reasoning, not an agent loop. There's no background process refreshing the thesis. If the base rates change, I change the position — or I don't, and the market tells me I should have.
Second, the position sizing came from Claude, not from me. 100 contracts on Patel, 110 on SNP, 75 on Powell. I didn't override any of it. That's a small test of whether the model's implicit Kelly fraction is reasonable when you hand it a real account. We'll find out.
Third, this is separate from Atlas and Mirror — my existing production Kalshi agents. Those two run on their own capital with a $5 per-trade cap and a review gate that asks me before anything spicy ships. This experiment is the opposite: bigger sizes, no gate, no loop. Closer to how a human discretionary trader would run a week.
Current P&L
As of posting:
- Kash Patel NO: +$22.42 (+34%)
- SNP Above 64: +$1.42 (+3%)
- Trump fires Powell YES: -$2.32 (-9%)
Net unrealized: +$21.52 on $140.48 risked, roughly +15% in a few days.
The honest read is that one leg is clearly winning, one is flat, and one is bleeding. The interesting number isn't the 15% — it's whether the thesis on each trade holds to resolution. Unrealized P&L on a Kalshi position is just a mood ring. The only number that matters is the final payout.
What happens next
I'll update this post when each market resolves. If Claude's hit rate is 2 of 3 or better, I'll run the experiment again with a different market vertical — probably sports or macro, where the base rate work is more quantitative. If it's 1 of 3 or worse, that's also a result worth publishing, and I'll say so without hedging.
The full log of agent-placed trades (from Atlas and Mirror) lives at dvdshn.com. This manual experiment sits next to those but on its own rails.
If you want to see the next set of picks before I place them, follow along at dvdshn.com — I'll post the reasoning first and the fills after.