Multi-Agent LLM System Steward for Autonomous Investing with Invariant Core Constraints on Self-Evolution

01 ABSTRACT

The authors present Steward, a multi-agent LLM system that draws the boundary of self-evolution in its type system, with nine role-specialised agents deciding jointly, a seven-layer shared memory, and quarterly reflector learning. In a five-month live deployment (102 trading days, 1,632 LLM calls), the system contained a 24.6% sector drawdown to a 0.3% loss; replay showed a single model given the same raw data traded more and lost 1.35 points against the market over five days where the joint design showed no significant loss. The authors argue the invariant core (order arithmetic, risk rules, discipline prompts) is key to safety and performance, but the live sample is only five months and replay/backtest results require further validation.

02 KEY FINDINGS

1. Nine role-specialised agents: four analysts emit typed reports, a Portfolio Manager sets target positions, and a protected Risk Manager audits.
2. A seven-layer shared memory assembled by code from the system's own record gives each decision position context, market narrative, calibration, and past decisions.
3. Code grades every trade and forecast nightly and writes grades into memory; a quarterly reflector proposes prompt learnings applied reversibly by a deterministic gate and verified by replay.
4. In five-month live deployment, a 24.6% sector drawdown was contained to a 0.3% loss; in replay, a single model lost 1.35 points against the market over five days while the joint design showed no significant loss.
5. The invariant core lowered maximum drawdown under all 50 random entry policies in a 2015-2026 backtest.