One ticker. Twelve specialists. Four phases.

Each analyst pulls live data from a single domain and writes what they see.

• market_analyst

  Reads last price, multi-day price action, volume vs the 60-day average, and the recent range. Pulls from yfinance by default, or from Alpaca when keys are configured.

• news_analyst

  Surfaces recent headlines about the ticker. Headlines flow into the transcript with source attribution so the reader can verify upstream.

• fundamental_analyst

  Asset-class-aware context — equities get earnings/valuation framing, crypto gets supply/macro framing. Never asserts numbers it cannot ground.

• sentiment_analyst

  Aggregates retail sentiment from StockTwits + the relevant subreddit (asset-class routed). Quantifies bullish vs bearish lean and notes message volume vs baseline.

Two researchers argue against each other; a manager arbitrates.

• bull_researcher

  Builds the strongest case to own the ticker. Synthesises the analysts, surfaces catalysts, and addresses the most likely counter-arguments before they are raised.

• bear_researcher

  Builds the strongest case against. Looks for confirmation bias in the bull thesis. Will name specific things that would falsify the trade.

• research_manager

  Reads both, weighs the evidence, and writes a synthesis that names which arguments are load-bearing and which are noise.

A single agent proposes a specific position — not just a direction.

• trader

  Takes the research synthesis and writes a concrete proposal: action (BUY/SELL/HOLD), conviction, target sizing as a percentage of a hypothetical portfolio, and an entry plan. This is the only phase that proposes a number.

Three risk seats stress-test the proposal from opposing angles, and a portfolio manager makes the final call.

• risk_aggressive

  Argues the proposed sizing is too small for the asymmetry on offer. Bias: upside under-pricing.

• risk_conservative

  Argues the position assumes the thesis is right and ignores tail risk. Bias: downside under-pricing.

• risk_neutral

  Holds the line between the two. Bias: the median path is most likely.

• portfolio_manager

  Reads all three risk seats plus the trader proposal and makes the final call. The decision card the user sees comes from this agent and carries a confidence score (0.0–1.0).