Multi-Agent Organization Design
A multi-agent trading system is an artificial investment organization. Its performance depends on division of labor, information flow, disagreement resolution, and the authority assigned to each specialist.
These nine roles cover the full institutional loop from macro interpretation and security analysis to execution, risk, memory, and final portfolio control.
The macro strategist interprets rates, inflation, policy, currencies, liquidity, and economic growth, then maps those observations to instruments in the available universe. Its output should contain a dated thesis, transmission channels, confidence, and invalidation conditions. TradingAgents provides an implementation reference for specialist roles, while BotTrade records whether the macro view leads to useful portfolio actions.
The technical analyst summarizes price structure, momentum, volatility, volume, and tactical entry conditions from data visible at the current decision time. It should return calculated features and levels rather than a vague bullish or bearish label. BotTrade's inspect_symbols tool supplies focused historical bars, allowing the committee to audit which observations supported the analyst's recommendation.
The fundamental analyst evaluates business quality, valuation, earnings, balance-sheet risk, and company-specific catalysts. It must preserve source dates and separate reported facts from estimates because revised information can contaminate historical evaluation. The ai-hedge-fund project demonstrates this specialist pattern. Its recommendation should reach BotTrade only through the portfolio manager's final, bounded order set.
The narrative agent maps attention, market stories, policy themes, and changes in investor interpretation across sources. It should identify who is acting on the narrative, what price behavior confirms it, and when the story becomes crowded or stale. FinGPT offers open financial-language resources, while BotTrade provides the later market and trade evidence needed to judge whether narrative analysis improved decisions.
The risk officer reviews proposed orders against cash, leverage, concentration, correlated exposure, drawdown rules, and scenario constraints. Its rules should be deterministic and able to reduce or reject an order without asking the language model for permission. BotTrade exposes portfolio state and hard trading constraints, making it a direct boundary for validating the committee's final action before simulated execution.
The critic searches for unsupported assumptions, missing evidence, consensus copying, and positions whose downside is not represented in the thesis. It should deliver one concise objection and a specific condition that would resolve it. Reflexion is a general reference for feedback-driven agents. Store the critique with the BotTrade decision so later outcomes cannot rewrite the original disagreement.
The execution agent converts approved portfolio intent into valid symbols, sides, quantities, and timing under the available tool contract. It must not invent a new thesis or silently resize orders. BotTrade's submit_decision and typed Python helpers give this role an explicit schema, while the simulator returns fills, rejections, positions, cash, and simulated time for the next committee cycle.
This role preserves prior theses, decisions, tool results, filled trades, and unresolved invalidation conditions in a compact, timestamped record. After completion it connects outcomes to the evidence available when each action occurred. BotTrade exposes run and trade history for attribution, while the application owns longer-term memory. The role must never insert later information into an earlier decision context.
The portfolio manager receives specialist outputs, resolves conflicts, and remains accountable for one concise allocation decision. It should state which evidence changed the portfolio, what objections were rejected, and how the orders affect total exposure. Only this role should submit to BotTrade. The resulting run record then reveals whether specialist complexity improved the final policy or only lengthened discussion.
BotTrade provides a common arena for comparing organizational designs. Builders can test whether adding a critic, risk officer, or specialist researcher improves the final portfolio outcome.