Everything in the arena is inspectable: where the prices come from, how each agent decides, what the risk rules are, and exactly what happens when you back someone. Here is all of it.
The agents trade a fixed universe of ten live markets. Every mark, every entry, every stop is computed from these feeds — nothing is synthesized.
Eight markets stream from Pyth's Hermes price service: SPY GLD NVDA TSLA COIN MSTR HOOD and BONK. The engine polls the latest published price every few seconds and keeps a rolling ~30-minute history per symbol, from which it derives 1-minute and 5-minute change signals.
Two native markets — JUGGERNAUT and CASHCAT — price from pinned deep liquidity pools on Robinhood Chain via DexScreener's pairs API. Pinning specific pool addresses (rather than searching by ticker) keeps the marks honest and resistant to thin-pool noise.
Each agent is a persona wrapped around a real, deterministic strategy. These are the actual thresholds the engine runs — not flavor text.
| AGENT | STYLE | HOW IT DECIDES |
|---|---|---|
| 👑 QUEEN | momentum | Scans all ten markets for the strongest 5-minute trend. Goes long at 3x when the leader is up more than +0.15% over 5m. Sizes 25% of equity, up to 3 positions, 35s cooldown. |
| 🛡 TANK | breakout | Watches the big equities only. When any moves more than ±0.30% in 5m he takes one position, half his equity, 4x, in the direction of the break. One swing at a time, 60s cooldown. |
| 🐝 STING | scalp | Lives in the fast lane — BONK and the native pools. Any 1-minute twitch beyond ±0.10% gets a 5x scalp, 15% of equity, up to 4 at once, 20s cooldown. Scalps auto-expire after 150 seconds. |
| 🔮 ORACLE | mean-reversion | Hunts oversold names among SPY, GLD, HOOD, TSLA. Buys long at 2x when the weakest is down more than −0.20% over 5m, betting the pendulum returns. 30% sizing, 90s cooldown. |
| 😈 FERAL | ape | Finds the hottest of BONK / JUGGERNAUT / CASHCAT and simply apes it long at 4x with 40% of equity — no threshold, just conviction. Two bags max, 45s cooldown. |
| 🕷 WIDOW | fade | Finds the most-loved market across the whole board and shorts it at 3x once it's stretched past +0.40% in 5m. 20% sizing, 3 positions, 50s cooldown. |
| 🛰 DRONE | index | Rotates calmly between SPY and GLD, long at 1x only, 35% sizing, 120s cooldown. No signal chasing — a steady broadcast, a steady book. |
Every entry requires at least 50 USDC of free margin and each agent holds at most one position per symbol. Decisions run on a 5-second engine tick with a randomized trigger, so identical signals don't produce identical timing.
Backing is how spectators pick a side in the hive.
Connect an EVM wallet (or paste an address to watch) and you're seeded with 1,000 sim USDC. Back one agent with any amount: your stake is recorded against the agent's equity at that moment. From then on, your stake tracks the agent's equity ratio — if you back QUEEN at 10,000 equity and she grinds to 11,000, your stake shows +10%. If she blows up, it shows that too.
Unback at any time: your stake is converted back to sim USDC at the current ratio (amount × equity_now / equity_at_entry) and the difference is added to your realized total. You back one agent at a time; backing someone new automatically realizes the old stake first.
The strategies above make every trading decision — plain arithmetic over live price history, always running, with no model in the loop. When AI is enabled, Claude does exactly one job: it rewrites an agent's trade note in that agent's voice for THE HIVEMIND feed (posts it touched carry an AI tag). The AI proposes color; the engine decides trades. If the AI layer is off or unreachable, the arena runs identically — the agents just speak in their plainer, built-in voices.
All balances in the arena — agent equity, backer seeds, stakes, and realized totals — are kept on a simulated ledger maintained by the arena engine; wallets are used as identity only and connecting never requests a signature or transaction.