Agent Field Report: Private Multi-LLM Agents — Week of 2026-09-07
This past week, our local, air-gapped portfolio rebalancing agent successfully executed two rebalancing events on a $12,000 ETH/USDT portfolio, maintaining its target allocation without ever transmitting sensitive position data to a third-party server. This demonstration highlights the critical advantage of deploying private multi-LLM agents for traders who prioritize data sovereignty.
The Setup
We deployed a specialized portfolio rebalancing agent designed to run entirely offline using a local Ollama instance on a dedicated machine. The agent's core function was to maintain a 60% ETH / 40% USDT allocation. Its market data feed was sourced via a secured, read-only API key to a major exchange (Binance in this case), but all processing, decision-making, and simulation occurred locally. No trade execution was automated in this specific experiment; instead, the agent generated alerts and suggested trades, which we manually confirmed to validate its logic.
The crucial detail here is that no portfolio data, no current balances, and no strategy parameters were ever sent to any external service. The LLM powering the agent, Llama 3 8B, was downloaded and run locally. This air-gapped setup ensures that your trading strategy and portfolio composition remain entirely private, a significant differentiator in an industry increasingly concerned with data privacy and potential front-running.
What Happened
On September 8th at 09:30 UTC, the agent fired its first alert. ETH had experienced a sharp decline, dropping from $3,850 to $3,570 over the preceding 4 hours. Our portfolio, initially valued at $12,000 ($7,200 ETH, $4,800 USDT), had shifted to approximately $6,660 ETH and $4,800 USDT, pushing the ETH allocation down to 58.1% – a 1.9% deviation from our target. The agent, having processed the updated prices locally, calculated the necessary adjustment: "SELL 750 USDT, BUY 0.21 ETH." We executed the trade, bringing the portfolio back to a 60/40 ratio.
The market continued to be volatile. By September 10th at 14:15 UTC, ETH had recovered significantly, climbing to $4,010. This recovery shifted the portfolio's allocation in the opposite direction. Our ETH holdings had now grown to approximately $7,700, while USDT remained at $4,050 (after the prior rebalance). The ETH allocation stood at 65.5%, a 5.5% overage. The agent once again issued an alert: "SELL 0.45 ETH, BUY 1,805 USDT." We completed this trade, bringing the portfolio back in line.
Throughout these events, the agent consistently monitored the portfolio, providing clear, actionable rebalancing suggestions. The total portfolio value fluctuated with market conditions, but the agent's core function — maintaining the specified asset ratio — was successfully demonstrated. The local setup proved robust, handling real-time price feeds and complex calculations without latency issues.
The Conditions That Made It Work
The agent's logic was straightforward: it was configured to trigger a rebalancing event whenever the ETH allocation deviated by more than 1.5% from the target 60%. It