All About AI

Find 100x Low Risk High Reward Polymarket Strategies With AI

By · · 5 min read

A few weeks ago my Polymarket bot hit a 100x trade — paid $1, redeemed $100. Then another at 50x, another at 48x. The wallet history was right there, but the why wasn't. So I did what I always do when something weird happens on-chain: I fed the proof-of-trade hashes into Claude Code and Codex and asked them to figure out what actually happened. This post walks through what I found, the two replication strategies the agents helped me design, and why this is the kind of workflow I keep coming back to — not because the money is huge, but because the loop between "find anomaly → analyze with an agent → ship a testable strategy" is incredibly fast.

Watch the video:

The strange trade

On the 5-minute BTC up/down market my bot somehow filled both sides — up and down — at one cent each, just as the window was closing. Total cost: $2. Down won, redeemed at $50. Up was a total loss at $0. Net: $48 from $2 in.

That should not be possible in a clean market. By the time a 5-minute window has ~1 second left, the implied probability is essentially settled — one side is near $0.95, the other near $0.05. Filling both legs at one cent means someone matched stale bids well after the outcome was already obvious. The question is: who, and why, and can it be reproduced?

Asking Codex what happened

The nice thing about Polymarket is everything settles on-chain, so the agent has receipts to work with. I copied the proof-of-trade screen, opened a fresh Codex session with --dangerously-skip-permissions, pasted the hashes in, and asked it to investigate. This is the same pattern as the autoresearch loop I ran on a hacker breach — let the agent chase wallets, transactions, and order-book traces in parallel while I do something else.

After a few minutes Codex came back with a clean reconstruction:

That last point is the key one. If you fill both up and down on the same window at one cent, you cannot lose. One side has to redeem. The trick is the fills are rare and depend entirely on a taker being willing to hit your stale low bids while the outcome is essentially known.

Brainstorming replication strategies

Once Codex understood the mechanism I asked it to brainstorm strategies for making this happen on purpose. It came back with a pile — two-sided cheap maker bids, post-close winner snipes, hybrid setups — but the two I actually wanted to test were simple enough to ship the same afternoon.

This is the part of the loop I keep coming back to in posts like running Claude Code agents on the predictions market: the agent is not picking the strategy. I am. The agent is doing the boring, high-volume part — reading the trade history, mapping the order-book mechanics, listing the plausible exploitations — and the resulting menu lets me ship two ideas in an hour instead of one in a day.

Strategy 1: preload future windows 24 hours ahead

Polymarket lists windows roughly 24 hours in advance. The order book on a fresh window is, predictably, empty. So the first idea is to walk into a new window the moment it appears and place $0.01 maker bids on both sides — up and down — across BTC, ETH, SOL, and XRP. Eight bids per window. They sit there for hours.

The economics are nice:

While recording I had the bot up to 100+ open orders sitting in future windows. None had filled yet, but that's expected — this is a long-tail strategy. If it works at all, you find out in days, not minutes.

Strategy 2: snipe the window switch, cancel after 120 seconds

The second strategy is more active. At the exact moment one 5-minute window closes and the next opens, fire $0.01 bids on both sides across all four crypto windows — eight bids again — and keep them live for exactly 120 seconds before cancelling.

Why 120 seconds? Because the price evolves through the window. At the open you're near 50/50, so a $0.01 bid is far out of the money and unlikely to be filled by mistake. By the end of the window prices have drifted toward 0.95/0.05, and a $0.01 resting order is a juicy target for anyone scraping the book. You want exposure to early-window weirdness without leaving stale bids around for the late-window scalpers. Cancel at the 2-minute mark and you skip the bad zone entirely.

This is the same parameter-tuning instinct I wrote about in building the original Polymarket bot: small bot, small bankroll, a few config knobs that move the entire risk profile. The whole point of doing this in Claude Code is being able to flip a number, restart, and watch.

Why this loop is the point

Neither of these strategies is going to make me rich, and that is fine. I run them with small money in part because I learn how prediction markets actually behave — naked positions, both-leg fills, maker rebates, settlement delays — and in part because they slot cleanly into the broader Claude Code passive-income setup I already have running.

What I actually got from this video was the workflow itself:

  1. Spot an anomaly in real data.
  2. Paste the receipts into an agent and ask it to explain.
  3. Once the mechanism is clear, ask the same agent for replication strategies.
  4. Pick the two most testable, ship them with cheap bids, let them run.

None of those four steps is hard on its own. The thing AI agents change is the time it takes to do all four end-to-end. A year ago this would have been a weekend of poking at Polyscan and writing Python. Today it is one afternoon, two strategies live, two parallel data collectors running. That compression is the real product.

I'll do a follow-up once either strategy has a meaningful number of fills to look at. If you want to talk strategies in the meantime, the AI_automata Discord is where I post live updates.

Resources