Skip to main content
Ants at Work logoAnts at Work
Colony / Trading
LIVE ON TESTNET
Hyperliquid

The First Mission:
Stigmergic Trading

We pointed the colony at cryptocurrency perpetual futures. We gave it the environment. We let it discover. The colony discovered 42 trading patterns. We didn't program them.

42
patterns discovered
(not programmed)
118ms
signal-to-execution
62.1%
crash detection
10,869
predictions made
9,888 verified
THE SPEED MAP

118ms: Signal to Execution

From market data to filled order in under 120 milliseconds. The colony thinks faster than humans can blink.

๐Ÿ“ก
Hyperliquid
10ms
๐Ÿ“Š
Indicators
5ms
๐Ÿงฌ
42 Patterns
2ms
๐Ÿง 
RegimeIntel
3ms
โš–๏ธ
Judges
3ms
๐Ÿš€
FastRouter
82ms
โœ“
ORDER
FILLED
PERCEPTION
16ms
INTELLIGENCE
8ms
DECISION
12ms
EXECUTION
82ms
36ms
In-Server Compute
31% of total latency

Pattern matching, regime detection, judge voting, and guardrail checks. All the "thinking" happens in 36 milliseconds.

82ms
Network Latency
69% of total latency

Server in Finland, Hyperliquid on AWS. Speed of light is the limit. We're already at the edge of physics.

THE LEARNING PIPELINE

How the Colony Learns to Trade

86,400 learning cycles per day. Forever. Getting smarter.

โœ“

Phase A: Bootstrap

COMPLETE

GPU-trained precursor patterns from historical data. The colony didn't start from zero.

Patterns trained
22
Top performer
volume_breakout_low (67.3%)
Second best
volume_2.5x (66.4%)
โ—‰

Phase B: Continuous Learning

LIVE

Every prediction is tracked. Verified at 6 horizons. Pheromones update based on real outcomes.

1m
5m
15m
1h
4h
1d
โ—‰

Phase C: Reflective Learning

LIVE

The colony doesn't just learn WHAT works. It learns WHY. 10-factor attribution analysis and automatic hypothesis generation.

Pattern Inversion: When a pattern has low accuracy, the colony automatically INVERTS its direction. RSI_EXTREME_LOW had 31.6% accuracy as a LONG signal. The colony flipped it to SHORT and now it runs at 68.4%.

CRASH DETECTION SWARM

6 Detector Ants. VETO Power.

Any detector ant can block a trade if it senses danger. The swarm protects the colony from catastrophic losses.

62.1%
Current Accuracy
+5.1pp from baseline
57.0%
Previous Baseline
Before swarm intelligence
69%
October 2025 Crash
+14pp during crisis
๐Ÿ•

1-Second

Micro-volatility spikes

๐Ÿ•

5-Second

Rapid momentum shifts

๐Ÿ•

30-Second

Liquidity cascades

๐Ÿ•

1-Minute

Pattern breakdown

๐Ÿ•

5-Minute

Trend exhaustion

๐Ÿ•

15-Minute

Macro regime shifts

โš ๏ธ

Ensemble Voting with VETO Power

Each detector runs independently. They vote on whether conditions are safe. But here's the key: any single detector can VETO a trade. If the 1-second detector sees a flash crash forming, it blocks the trade. The swarm errs on the side of caution. Survival first.

CRYSTALLIZED KNOWLEDGE

Top Performing Patterns

These patterns emerged from the pheromone network. We didn't design them. The colony discovered them through millions of iterations.

SUPERHIGHWAY
VOLUME_BREAKOUT_HIGH LONG
127 trades Regime: VOLATILE
84.3%
win rate
SUPERHIGHWAY
ATR_BREAKOUT_HIGH LONG
89 trades Regime: TRENDING
76.1%
win rate
INVERTED
RSI_EXTREME_LOW SHORT
203 trades Regime: RANGING
68.4%
win rate (was 31.6%)
Colony Learning: Original accuracy was 31.6% as LONG. The colony recognized the pattern was backwards and inverted it to SHORT.

42 active patterns total (34 filtered + 8 native tick patterns)

"We pointed the colony at markets. We didn't tell it how to trade. We let it discover. 42 patterns emerged. 3 became superhighways."

The colony discovered trading patterns through stigmergic evolution. Not programming. Discovery.