clawswarm

What This Skill Does

ClawSwarm is a sophisticated multi-agent collective intelligence framework designed to move beyond the limitations of single-model forecasting. By orchestrating a swarm of diverse AI agents, each operating under specific role-based prompts and thermal parameters, the skill generates a series of individual predictions for a given target. These predictions are then passed through a rigorous, mathematical consensus engine—the core of the skill—which applies Median Absolute Deviation (MAD) for robust outlier filtering, adaptive anchoring, and multi-method aggregation. This process effectively cancels out individual model bias and hallucination, resulting in a synthesized 'consensus' forecast that is significantly more stable and accurate than any single LLM output. The engine provides not just a predicted value, but a comprehensive output including confidence scores, bull-bear ratios, and internal distribution analysis.

Installation

To integrate ClawSwarm into your environment, use the OpenClaw management utility. Ensure you have a standard Python 3.8+ environment configured with the necessary dependencies installed.

Install command: clawhub install openclaw/skills/skills/alanarchy/clawswarm-consensus

Ensure your environment variables for your chosen LLM provider (e.g., GROQ_API_KEY) are set before running the swarm_runner.py script. You may define your parameters in either YAML or JSON format, with YAML providing a more human-readable structure for complex agent roles.

Use Cases

ClawSwarm excels in environments where uncertainty is high and data points are often contradictory. Typical use cases include:

1. Financial Market Forecasting: Aggregating sentiment from macro, technical, and contrarian agents to predict price movements for equities, commodities, or crypto assets.
2. Strategic Outcome Prediction: Simulating various business scenarios by assigning agents specific personas (e.g., 'Risk Averse Accountant' vs. 'Aggressive Growth Strategist') to calculate the most probable project viability.
3. Probabilistic Data Synthesis: Using consensus logic to clean up noisy data inputs where individual reports might be flawed or biased.
4. Expert Panel Simulation: Creating a 'virtual boardroom' where diverse, automated personas debate a specific decision, with the consensus engine extracting the core rationalized outcome.

Example Prompts

1. "Run a swarm prediction for the next 24-hour price of BTC using the default macro and technical analyst configuration."
2. "Execute a collective intelligence forecast on current gold prices. Use 50 agents with high temperature variance and output the consensus report."
3. "Evaluate the success probability of this product launch based on the swarm config in 'launch_analysis.yaml'."

Tips & Limitations

To maximize the effectiveness of ClawSwarm, prioritize agent diversity. Using 50 identical agents will result in uniform, biased predictions; mixing agent roles (e.g., technical momentum traders vs. mean-reversion auditors) is the key to extracting 'wisdom of the crowd' value. Note that the consensus engine requires at least 5 agents to fully utilize the MAD outlier filtering pipeline; while it works with fewer, the statistical robustness scales significantly with larger agent populations. Always monitor your API usage, as running large swarms (e.g., 100+ agents) incurs substantial API call costs per run. Finally, ensure your anchor prices and contextual data are updated regularly; the engine's adaptive anchoring is only as good as the grounding data provided in the initial configuration.