Anthropic just open-sourced the exact playbook their engineers use

It explains how to build production AI agents:
multi-agent systems, memory layers, orchestrators, context engineering.
Most people scrolled past it like another random GitHub repo.
I didn’t.
In one day I built 3 agents with a single mission:
find Polymarket wallets that make money not by predicting events, but by exploiting structure.
They discovered 3 addresses.
Each used the same principle, just on different markets.
Wallet #1 - Elon Musk tweet markets
It bought all 27 possible outcomes at ~3¢ each.
Result:
Won 7 out of 27 outcomes → still profitable.
One series alone made $2,270.
Total PnL: $130,000
Wallet #2 - New York temperature markets
Bought all 9 outcomes at ~6¢ each.
Result:
Won 1 out of 9 → still profitable.
$1,000 invested → $250 profit
Total PnL: $75,000
At first I couldn’t understand the math.
Then it clicked.
If a market has 10 outcomes priced at 5¢ each, buying all of them costs 50¢ total.
But the winning outcome pays $1.00.
You profit no matter which outcome happens.
You don’t need predictions.
You only need markets where the sum of all probabilities < 100¢.
This isn’t betting.
It’s probability arbitrage.
Then I added a 4th agent.
Built on the LLM-as-Judge principle - an agent that doesn’t search, it evaluates.
It stress-tested all three wallets for:
• drawdowns
• stability
• strategy resilience
It chose this wallet:

Wallet #3 — Trump Truth Social post markets
Strategy:
Buy all possible posting ranges at ~10¢.
Example result:
Win 3 out of 6 outcomes → still profitable.
Daily profit: $674 – $1,100
No predictions.
No news.
No luck.
The judge agent picked this wallet not because it earned the most, but because the strategy is the most stable.
Weather can swing wildly.
Tweet counts are unpredictable.
But Trump posting frequency is statistically consistent — because Trump behaves like Trump.
So I connected a copy-trading bot and started mirroring every trade:

Now I don’t:
• analyze markets
• calculate probabilities
• follow the news
Three agents found the strategy.
The fourth selected the best one.
The bot simply copies it.
Anthropic gave everyone the same tools.
The only question is what you build with them.