Agent Field Report: Private Multi-LLM Agents — Week of 2026-09-06
This week, we put a core tenet of secure trading to the test: complete data isolation. Our private multi-LLM agents successfully executed a critical portfolio rebalance for a BTC/ETH pair, entirely off-grid. This wasn't just a technical exercise; it was a demonstration of how traders can protect sensitive position data from cloud exposure, ensuring their strategy remains their own.
The Setup
We deployed a dedicated portfolio rebalancing agent with a single objective: maintain a 60/40 Bitcoin (BTC) to Ethereum (ETH) allocation within a $100,000 hypothetical portfolio. The critical difference from typical cloud-based solutions? This agent ran entirely on a local Ollama instance on an air-gapped machine. There were no external API calls for LLM inferences, no third-party database storing portfolio states, and absolutely no data leaving our controlled environment. We utilized Llama 3 8B for initial market sentiment filtering and Mixtral 8x7B for the complex rebalancing calculations and trade order generation, both running concurrently on the local hardware.
The system was configured to pull real-time price data from a self-hosted data feed, ensuring that even market data acquisition was isolated from commercial API providers. The initial portfolio stood at $60,000 in BTC (approx. 0.90 BTC at $66,666) and $40,000 in ETH (approx. 11.42 ETH at $3,500). Our goal was to prove that sophisticated algorithmic trading decisions, historically reliant on cloud infrastructure, could be made with absolute privacy, addressing the growing concern among professional traders about their strategies becoming part of a third-party data lake.
What Happened
The agent began monitoring the BTC/ETH pair at 00:00 UTC on September 7th. For the first 36 hours, market movements kept the portfolio within our acceptable 2% deviation threshold. However, at 03:17 UTC on September 8th, a sharp downturn in BTC, coupled with a modest rise in ETH, triggered the agent. Specifically, BTC dropped from $66,666 to $64,800 (-2.79%), while ETH climbed from $3,500 to $3,580 (+2.28%). This shift pushed the portfolio allocation to approximately 58.7% BTC and 41.3% ETH, exceeding our pre-defined 2% deviation threshold from the 60/40 target.
Upon activation, the Mixtral 8x7B LLM, fed the current portfolio state and market prices, calculated the necessary adjustments. It determined that to restore the 60/40 balance, the agent needed to sell 0.045 BTC and buy 0.82 ETH. These calculated orders were then forwarded to a simulated exchange API (again, running locally) for execution. Within 30 seconds of the trigger, the rebalance was