Walsh is a multi‑agent stock research system that routes each ticker through four specialized agents in parallel—Fundamental, Technical, Sentiment, and Macro—aggregating their theses via a confidence‑weighted Portfolio Manager. The final call is gated by a Rule‑Based Risk Manager that can veto or downgrade trades before execution. Backtesting on AAPL, MSFT, and NVDA from Jan 2023 to Dec 2023 with $100k initial capital yielded a total return of 75.5% versus 119.8% for simple buy‑and‑hold, underscoring the conservative bias of the risk gate. In Oct 2023 the pipeline exited all positionsěli when momentum signals broke down, holding cash for the month and missing the subsequent NVDA rally—an intended behaviour that highlights the cost of a high‑uncertainty threshold in trending markets.

The system’s architecture features an Orchestrator that gathers agent outputs asynchronously, a Portfolio Manager that synthesizes signals with confidence weighting, and a Risk Manager that enforces hard veto rules (confidence < 45% or agent spread ≥ 4) and soft downgrades (confidence 45–65%). Known limitations include underperformance due to a volatility dampener (vol_cap = 40%) and mean‑reversion agents that issue sell signals on 52‑week highs. Adversarial testing identified overconfidence in 5 of 18 cases, revealing the need for better signal calibration and price validation.

Future updates plan to replace rule‑based agents with LLM‑backed counterparts, add cost and latency metrics, and refine the risk thresholds to adapt to trend versus mean‑reversion regimes.