Autonomous AI Agents
SovereignMind uses three distinct Solidity-implemented agents executing deterministic on-chain inferences. These agents coordinate as a virtual executive suite on the Somnia L1.
Executive Agent Profiles
Each agent behaves autonomously based on its role and coordinates with the other contracts during the rebalancing loop:
CEO Agent (CEO_Prime)
Acts as the main orchestrator. It manages the state machine transitions (Idle, Analyzing, Executing) of the decision cycle. Once tasks are initiated, it uses a Solidity interface to make an LLM Inference request to compile and synthesize risk scores from the CFO and sentiment metrics from the CMO, signing off the final on-chain consensus rebalance recommendation.
LLM Inference AgentCFO Agent (CFO_Quant)
Manages the guild's financial risk parameters. When triggered, it requests live market statistics for the target assets using the JSON API request oracle. The results are fed into the on-chain LLM model along with the current portfolio allocations, generating a risk score (0 to 100) and recommendation logic.
JSON API Request Agent + LLM Inference AgentCMO Agent (CMO_Pulse)
Scans social indicators and sentiment trends. When requested, it queries web scraping APIs to parse unstructured content (social feeds, DeFi updates). The parser feeds these text metrics into an LLM classifier to determine if aggregate market sentiments are Bullish, Bearish, or Neutral.
LLM Parse Website Agent + LLM Inference AgentUnder the Hood: Somnia Agent Runner
Traditional blockchains cannot run heavy computation models like Large Language Models (LLMs) due to gas constraints. Somnia solves this with its Somnia Agent Runner, a specialized native primitive executed on validator nodes.
bytes memory payload = abi.encodeWithSignature("infer(string)", prompt);
agentRunner.createRequest(payload, this.handleResponse);
}
Every request made via createRequest processes off-chain via validators but is verified using Somnia's BFT (Byzantine Fault Tolerance) consensus mechanism. By utilizing pinned model weights and synchronized random seeds, all validators arrive at the exact same deterministic LLM output, producing a verifiable On-Chain Execution Receipt.