Agent Field Report: Private Multi-LLM Agents — Week of 2026-09-04
Last week, our experimental rebalancing agent, a prime example of private multi-LLM agents, running entirely on a local Ollama instance, executed a critical portfolio adjustment for a BTC/ETH pair. This move prevented a 7% drawdown on the ETH leg relative to BTC, capturing upside from a subsequent recovery — all without sending a single byte of trade data or position information to a third-party server. This wasn't just a successful trade; it was a proof of concept for truly air-gapped, privacy-first AI in trading.
The Setup
We deployed this agent on September 1, 2026, on a dedicated Mac Studio M2 Max. The goal was to maintain a 60/40 BTC/ETH portfolio split. This wasn't a cloud deployment; the entire operation — market data ingestion, LLM inference, and trade execution logic — ran locally. We used Ollama to host two distinct models: Llama 3 (8B Instruct) for macro-level market context and sentiment analysis, and Mistral 7B for precise calculation and execution instructions. Both models were pre-loaded, requiring no external API calls, making the system fully air-gapped.
The agent was configured to monitor 1-hour candlesticks for both BTC/USD and ETH/USD, with a rebalancing threshold set at 5%. If the portfolio's deviation from the 60/40 target exceeded 5% in either direction, the LLMs would analyze the situation before recommending a rebalance. Our total simulated