Hive

13 compiled Rust subsystems coordinate a self-evolving fleet of trading agents.

The Hive is the brain of the Horizon swarm. It takes independent trading strategies and turns them into a self-organizing system that evolves, learns, and improves on its own. No proprietary algorithms are exposed.

All 13 core subsystems are compiled to Rust via PyO3. When you install the SDK, you get an opaque .so binary. No source code ships in the Python package.

Architecture

Compiled Core

13 subsystems compiled to native binary via PyO3/maturin. Pheromone fields, evolutionary engines, behavioral cloning, causal discovery. All opaque Rust code.

Python Orchestration

HiveController is a thin Python wrapper that wires the 13 Rust subsystems together. Configuration, persistence (SQLite), WebSocket server, and the oversight loop stay in Python.

Progressive Autonomy

Five trust modes from observe (watch only) to skynet (full self-modification). The system earns autonomy through demonstrated performance. Each graduation requires statistical evidence.

No IPC

Direct PyO3 function calls between Python and Rust. No sockets, no serialization overhead, no process boundaries. A Python call to pheromone.deposit() executes compiled Rust in the same process.

Quick Start

from horizon.hive import HiveController, HiveConfig, AutonomyMode

config = HiveConfig(
    total_capital=100_000,
    max_agents=50,
    initial_mode=AutonomyMode.Observe,  # Start conservative
    host="0.0.0.0",
    port=8780,
    auth_token="your-secret-token",
)

hive = HiveController(config)

# Spawn agents from 9 archetype templates
agent = hive.spawn_agent("momentum", markets=["BTC-YES"], capital=5000.0)

# Full status across all subsystems
status = hive.status()

# Start the oversight loop + WebSocket server (blocking)
hive.run()

Remote connections:

from horizon.hive import HiveClient

client = HiveClient("ws://hive-server:8780", token="your-secret-token")
client.connect()
print(client.status())
client.spawn(template="arbitrage", markets=["BTC-YES", "ETH-YES"])

13 Subsystems

Every subsystem listed below runs as compiled Rust. The Python layer is thin wrappers only.

Coordination

Subsystem	Purpose
SwarmCoordinator	Fleet management. Tick loop: fitness scoring, selection pressure (cull underperformers), promotion pipeline (paper to shadow to live), capital rebalancing via fitness-weighted allocation.
PheromoneField	9-channel stigmergic communication. Agents deposit pheromone (bullish, bearish, volatility, crowding, opportunity, toxicity, liquidity, regime_trend, regime_mr). Evaporation + diffusion create emergent market signals without direct messaging.
ConsensusEngine	3-layer collective intelligence. IC-weighted signal combination, internal prediction markets (agents bet on outcomes), and Dempster-Shafer belief aggregation for conflicting views.

Evolution

Subsystem	Purpose
AgentFactory	9 archetype templates (momentum, mean_reversion, market_maker, arbitrage, volatility, event_driven, macro, copy, liquidation) plus 5 additional spawn sources: evolution, clone, user, imprint, opportunistic.
MAPElitesArchive	300-cell quality-diversity grid (5x5x3x4) over behavioral descriptors: holding period, directional bias, volatility preference, asset focus. Keeps the swarm diverse by construction.
EvolutionEngine	NSGA-II multi-objective optimization with island model. Wraps the existing src/evolution.rs Rust module. Populations evolve across 5 islands with periodic migration.

Intelligence

Subsystem	Purpose
CriticalityMonitor	15-indicator systemic risk assessment. Correlation convergence, contagion score, Hurst exponent, two-scale volatility, market entropy, transfer entropy, VPIN, liquidity, spread widening, Amihud ratio, BOCPD, CUSUM. Triggers avalanche reserve deployment on criticality.
CausalEngine	Granger causality + transfer entropy + lead-lag detection across all tracked markets. Builds a causal graph that detects when relationships break. Early warning for regime shifts.
SelfImpactModel	Almgren-Chriss calibrated impact estimation. OLS regression on observed slippage. Gates orders that would move the market more than the expected alpha. Internal crossing for opposite-side fills.
ImprintSystem	Behavioral cloning via 3-layer MLP (16 to 64 to 32 to 9 actions). Learns from operator actions with DAgger-style corrections. MaxEnt IRL reward model. Learns your trading style and can suggest or replicate it.

Safety

Subsystem	Purpose
AutonomyController	5-mode progressive autonomy state machine. Each transition requires statistical evidence (acceptance rate, Sharpe comparison, incident-free period). Observe, Suggest, Supervised, Autonomous, Skynet.
KillChain	6-phase automated incident response: detect, contain, preserve, diagnose, remediate, learn. Python callbacks for each phase. Genome blacklisting prevents failed strategies from re-evolving.
MetamorphicExecutor	Anti-fingerprint execution. Slices orders into randomized child orders with jittered size, exponential timing, beta-distributed aggression, round-robin venues, and Gaussian price offsets. Optional decoy orders.
ReputationSystem	Bayesian trust scoring with Beta(alpha, beta) posteriors per agent. Behavioral fingerprinting detects style drift. Reputation weights capital allocation and pheromone deposits.

Autonomy Modes

The Hive doesn’t start autonomous. It earns trust through demonstrated performance.

Mode	Behavior	Graduation Criteria
Observe	Watch only. No actions, no suggestions.	100+ imprint observations
Suggest	Suggest actions, never execute. Human reviews.	70%+ acceptance rate over 14 days
Supervised	Act with human approval for non-routine decisions. Auto-approve routine actions.	80%+ approval rate over 28 days, Sharpe > threshold
Autonomous	Independent within guardrails. All actions allowed.	6+ months profitable, Hive Sharpe > user Sharpe, 0 critical incidents in 90 days
Skynet	Full autonomy with self-modifying guardrails.	Manual promotion only

# Check current mode
mode = hive.autonomy.status()["mode"]

# Force mode change (bypasses graduation)
hive.autonomy.set_mode(AutonomyMode.Supervised)

# Pending approvals in supervised mode
pending = hive.autonomy.pending()
hive.autonomy.approve(pending[0]["id"])

Pheromone Field

Agents communicate indirectly by depositing pheromone on markets. No direct messaging. Emergent coordination through environmental modification. Same mechanism ant colonies use.

from horizon._horizon import PheromoneChannel

# Agent deposits bullish signal on BTC, weighted by its fitness
hive.pheromone.deposit("BTC-YES", PheromoneChannel.Bullish, amount=0.8, agent_fitness=1.5)

# Another agent senses the field before trading
readings = hive.pheromone.sense("BTC-YES")
print(f"Bullish: {readings.bullish}, Crowding: {readings.crowding}")

# Pheromone evaporates over time (old signals fade)
# Diffuses to related markets (BTC signal bleeds to ETH)
# Crowding channel auto-updates based on agent count
hive.pheromone.tick(dt=1.0)

# Full field snapshot
snapshot = hive.pheromone.snapshot()  # market_id -> channel -> value

Nine channels:

Channel	What It Signals
Bullish / Bearish	Directional conviction from agents actively trading
Volatility	Agents detecting unusual price movement
Crowding	How many agents are on the same market (auto-computed)
Opportunity	Untapped edge detected but not yet exploited
Toxicity	Adverse selection or market manipulation detected
Liquidity	Available depth for execution
RegimeTrend / RegimeMR	Current market regime classification

Kill Chain

Automated incident response. Six phases in sequence with Python callbacks at each step.

from horizon._horizon import Severity

# Register response handlers
hive.kill_chain.on_contain(lambda incident: pause_affected_agents(incident))
hive.kill_chain.on_preserve(lambda incident: snapshot_full_state(incident))
hive.kill_chain.on_diagnose(lambda incident: identify_root_cause(incident))
hive.kill_chain.on_remediate(lambda incident: kill_and_blacklist(incident))
hive.kill_chain.on_learn(lambda incident: record_for_evolution(incident))

# Trigger manually or automatically via criticality monitor
hive.kill_chain.trigger(
    Severity.Sev1Critical,
    "portfolio_drawdown_15pct",
    ["agent-001", "agent-002"],
    pnl_impact=-15000.0,
)

Configuration

HiveConfig is a Python dataclass with 42 fields. Key settings:

Setting	Default	Purpose
total_capital	100,000	Total AUM across all agents
max_agents	50	Maximum concurrent agents
initial_mode	Observe	Starting autonomy mode
max_drawdown_pct	0.20	Portfolio-wide drawdown limit
max_capital_per_agent_pct	0.15	No single agent gets more than 15%
pheromone_tick_interval	1.0s	Pheromone evaporation/diffusion rate
criticality_check_interval	60s	How often to run the 15-indicator check
avalanche_reserve_pct	0.15	Capital held in reserve for crises
dream_enabled	True	Offline accelerated evolution between sessions
imprint_enabled	True	Learn from operator behavior
metamorphic_enabled	True	Anti-fingerprint execution
host / port	0.0.0.0:8780	WebSocket server for remote connections
auth_token	None	Required for remote client auth

CLI

25 commands under horizon hive:

# Status
horizon hive status
horizon hive agents
horizon hive agent <agent-id>

# Lifecycle
horizon hive spawn --template momentum --markets BTC-YES --capital 5000
horizon hive kill <agent-id> --reason "negative sharpe"
horizon hive pause <agent-id>
horizon hive resume <agent-id>
horizon hive scale <agent-id> --capital 10000

# Autonomy
horizon hive mode                    # Current mode
horizon hive pending                 # Decisions awaiting approval
horizon hive approve <decision-id>
horizon hive reject <decision-id>

# Intelligence
horizon hive evolve --generations 100
horizon hive dream --generations 10000
horizon hive pheromone BTC-YES
horizon hive reputation
horizon hive consensus BTC-YES
horizon hive criticality
horizon hive incidents

# Emergency
horizon hive kill-switch             # Pause everything immediately

# Remote
horizon hive connect ws://server:8780 --token secret
horizon hive server --port 8780      # Start WebSocket server

MCP Server

22 tools for LLM-driven Hive management via horizon.mcp_hive:

# Start the MCP server
python -m horizon.mcp_hive

# Or with auth
HIVE_MCP_TOKEN=secret python -m horizon.mcp_hive

Tools include: hive_status, hive_spawn, hive_kill, hive_evolve, hive_dream, hive_pheromone, hive_consensus, hive_criticality, hive_incidents, hive_kill_switch, and more. Full tool definitions are in python/horizon/mcp_hive.py.

What the Hive Cannot Do

Cannot Override Rust Risk

The Rust engine’s 8-point risk pipeline sits below the Hive. Every order goes through compiled checks regardless of what the Hive decides.

Cannot Skip Kill Chain

Once a kill chain triggers, all 6 phases execute. The Hive can’t suppress an incident or skip remediation. Genome blacklisting is permanent.

Cannot Reverse Graduation

Autonomy modes only go forward. The system can’t demote itself from Autonomous back to Supervised without human intervention.

Cannot Exceed Capital Limits

Total allocation across all agents can’t exceed total_capital. The coordinator enforces this in compiled Rust. No Python code can bypass it.

IP Protection

The proprietary algorithms (pheromone fields, evolutionary engines, behavioral cloning, causal discovery, impact modeling, metamorphic execution) are all compiled to native binary. When you publish the SDK:

• Users get _horizon.cpython-312-darwin.so (or equivalent for their platform)
• No .py files contain algorithm implementations
• The Python hive/ directory contains only configuration, persistence, and thin wrappers
• Decompiling the .so produces assembly, not readable algorithms