ClawSwarm#

Multi-agent collective intelligence framework. Run N agents with different analytical perspectives, aggregate predictions through a statistical consensus engine.

Quick Start#

1. Create a config file#

yaml

target:
  name: "Gold"
  current_price: 5023.1
  unit: "USD/troy oz"
  context: "RSI: 40.8 | MA5: 5084 | MA10: 5120"

agents:
  - role: "Macro analyst focusing on geopolitical risk"
    count: 50
    temperature_range: [0.4, 0.7]
  - role: "Technical RSI/MACD momentum trader"
    count: 30
    temperature_range: [0.45, 0.6]
  - role: "Mean reversion auditor"
    count: 20
    temperature_range: [0.35, 0.55]

api:
  provider: groq
  model: llama-3.3-70b-versatile
  api_key_env: GROQ_API_KEY
  delay_ms: 1200

consensus:
  max_deviation: 0.15

2. Run the swarm#

bash

python3 scripts/swarm_runner.py --config swarm.yaml

Output: JSON with final_price, median_price, confidence, bull_ratio, and all individual predictions.

3. Run consensus standalone#

Pipe any predictions array to the consensus engine:

bash

echo '{"predictions":[{"price":100.5,"confidence":70},{"price":99.8,"confidence":60}],"anchor_price":100.0}' \
  | python3 scripts/consensus.py

Architecture#

text

Config (YAML/JSON)
  ↓
Swarm Runner (swarm_runner.py)
  ├─ Agent 1 → LLM API → prediction
  ├─ Agent 2 → LLM API → prediction
  ├─ ...
  └─ Agent N → LLM API → prediction
  ↓
Consensus Engine (consensus.py)
  ├─ Bias correction
  ├─ MAD outlier filtering
  ├─ Anchor-distance filtering
  ├─ Multi-method aggregation (weighted 40% + median 35% + trimmed mean 25%)
  ├─ Adaptive anchoring (dispersion → anchor strength)
  └─ Clamping
  ↓
Final consensus prediction + confidence + bull/bear ratio

Key Concepts#

Agent diversity: Each agent gets a different role prompt and temperature. More diversity = better consensus.

Consensus engine: Not a simple average. Uses MAD (Median Absolute Deviation) to filter outliers, adaptive anchoring to stabilize results when predictions are dispersed, and multi-method aggregation for robustness.

1 agent or 1000: Works with any count. Single agent bypasses consensus. 5+ agents get full pipeline.

Config Reference#

See references/config-reference.md for full field documentation and example configs.

Scripts#

ScriptPurposescripts/swarm_runner.pyOrchestrate multi-agent predictionsscripts/consensus.pyStandalone consensus engine (pipe JSON in)

Dependencies#

• Python 3.8+
• numpy (for consensus engine)
• requests or urllib (for API calls)
• pyyaml (optional, for YAML configs; JSON always works)