Agent Field Report: Private Multi-LLM Agents — Week of 2026-04-19
This past week, our team at Assistant Hub ran an experiment that underscores a critical shift in AI-driven trading: complete data sovereignty. We successfully deployed a portfolio rebalancing agent that operated entirely on a local server via Ollama. No cloud services, no external API calls for its decision-making, and fully air-gapped from the broader internet. This setup demonstrates a practical application of private multi-LLM agents for traders who demand absolute privacy for their strategies and positions.
The agent’s core function was to maintain a specific asset allocation without exposing any portfolio data to third parties. We observed its first live trade execution, buying 0.05 ETH, triggered by a predefined deviation. The entire process, from market data ingestion to trade signal generation, happened within our controlled environment.
The Setup
We deployed a specialized portfolio rebalancing agent designed to manage an ETH/USDT allocation. The primary objective was to maintain a 60% ETH, 40% USDT split within a test portfolio. This agent ran on a dedicated Mini PC equipped with 32GB RAM and an NVIDIA RTX 4060 GPU, hosting Ollama. The machine itself was isolated, receiving its market data feed via a one-way data diode, meaning no outbound network traffic was possible from the agent's operational core.
Our initial portfolio consisted of 0.1 ETH and 200 USDT. The agent utilized a fine-tuned Llama 3 model for strategic decision-making and a Mistral 7B model for market data interpretation – both running concurrently on Ollama. The Llama 3 model was specifically instructed on rebalancing logic, while Mistral assisted in parsing raw market tick data into a format Llama 3 could easily consume for portfolio valuation. This dual-model approach, powered by local Ollama, allowed for robust analytical capabilities within a fully private environment.
What Happened
On 2026-04-21 at 09:15 UTC, the agent detected a significant deviation from its target allocation. ETH, which had been trading stably around $2,000, experienced a rapid decline to $1,850 over 15 minutes. This price action pushed the ETH portion of our portfolio down to 58.7% of the total value, falling below our predefined 58% threshold.
The agent, operating entirely offline, immediately identified the imbalance. At 09:17 UTC, the Llama 3 model, after receiving the updated portfolio valuation from the Mistral model, generated a buy signal. It calculated the necessary purchase amount to bring the portfolio back into target range, issuing an instruction to acquire 0.05 ETH against USDT. This instruction was then relayed via a secure, local message queue to an authenticated trading bot running on a separate, internet-connected machine (the only component with API access).
The buy order for 0.05 ETH was executed at an average price of $1,860. At 09:20 UTC, the agent confirmed the trade and updated its internal portfolio record. The new portfolio composition was approximately 0.15 ETH and 107 USDT, restoring the ETH allocation to 61.2% – comfortably within our target range. We received a secure, end-to-end encrypted notification