Building a cross-platform arbitrage agent for Polymarket & Kalshi: what I learned

Six months ago I started noticing something: the same binary event would have different prices on Polymarket and Kalshi. Buy both sides for less than $1.00 total, guaranteed $1.00 payout regardless of outcome.

The problem: windows last 30-200 seconds. Manual trading catches maybe 5% of them.

So I built an automated agent.

How it works technically:

• Python async engine polls both order books continuously

• Detects net edge after fees (both platforms charge ~1-2%)

• Kelly criterion sizes the position relative to bankroll

• Fires both legs in parallel via official APIs

• Circuit breaker: if one leg fails, the other unwinds in <400ms

• Telegram notification on every execution

What surprised me:

• Spreads appear most often around breaking news (elections, Fed decisions, sports events)

• Leg risk is the real risk, not “is this arbitrage real”

• Fee calculation has to be exact or you trade yourself negative

Where it is now:

Live on waitlist at https://arbitrage-agent.com/, still early, building the subscriber base before opening access.

Happy to answer questions about the engine, the execution model, or the maths.