Systems Architecture Review
Autonomous trading is not a single design pattern. The field spans compact decision loops, specialist committees, research swarms, reflexive critics, and hierarchical portfolio systems.
Architecture determines how information becomes a position. These twelve patterns organize the main choices around specialization, memory, debate, execution, and risk. Each entry includes a paper, repository, or implementation reference, plus a practical test for connecting the design to BotTrade's market simulator.
One model alternates between inspecting state, selecting a tool, reading the result, and submitting a portfolio action. This is the best starting architecture because every failure remains attributable to one policy and one context window. Implement it with BotTrade's scan, inspect, and decision tools, then examine whether the agent gathers only the evidence it needs before trading.
A research phase produces a structured thesis, catalysts, invalidation conditions, and candidate symbols before the execution phase sees the live portfolio state. The separation prevents exploratory notes from becoming accidental orders. The ai-hedge-fund repository provides a concrete reference for specialist financial analysis, while BotTrade can supply the subsequent scenario, simulated fills, and performance record.
Specialists analyze fundamentals, technicals, sentiment, or macro conditions, then a portfolio manager resolves conflicts and creates the final order set. The design is useful when evidence sources require different prompts or tools, but token cost and correlated opinions can grow quickly. TradingAgents is an implementation reference; a BotTrade run can reveal whether debate changes positions or merely produces longer rationales.
A top-level allocator sets cash reserves, asset budgets, leverage, and risk limits while lower-level policies choose instruments and timing. This structure fits multi-asset systems because local agents cannot silently consume the entire portfolio budget. FinRL offers examples of financial decision environments, while BotTrade provides an agent-facing simulator where the hierarchy's final orders and portfolio consequences can be inspected.
The trader proposes an action and a critic searches for unsupported assumptions, missing risk, contradictory orders, or an absent invalidation condition. The trader may revise once before submission. Reflexion is a useful general reference for feedback-driven agents. In a trading implementation, cap the review loop and measure whether criticism reduces drawdown or invalid orders rather than rewarding elaborate but inactive analysis.
The agent retrieves earlier theses, decisions, and outcomes that resemble the current state, then states why the memory is relevant before using it. FinMem is a financial-agent reference for layered memory. Guard against hindsight leakage and stale analogies by storing only information available at each simulated time. BotTrade's sequential observations and trade record make those temporal boundaries testable.
A classifier assigns the current observation to a market state and routes control to a specialized policy, prompt, or risk budget. The router should expose its features and confidence because a wrong classification affects every downstream action. Qlib provides an open quantitative research foundation for regime features and models; BotTrade supplies historical scenarios where routing decisions lead to simulated portfolio outcomes.
Separate workers own market scanning, symbol analysis, external research, and portfolio accounting while an orchestrator controls their order and context. LangGraph is a concrete framework for stateful graph workflows. Keep the final trade authority in one node, pass typed results instead of prose transcripts, and use BotTrade's MCP server as the execution boundary so tool ownership remains clear.
This system activates around earnings, policy decisions, protocol events, or other scheduled and unscheduled catalysts. It maps the event to affected instruments, defines a time horizon, and exits when the thesis expires. FinGPT provides open financial-language resources for event interpretation. BotTrade can test the resulting policy on historical bars without pretending the language model already knew later prices.
Several agents form decisions independently, and an aggregator combines their directions, sizes, and confidence into one bounded portfolio action. Independence matters because copies of one prompt tend to repeat one failure. Qlib's ensemble tooling offers a quantitative reference. Record every component vote alongside the submitted BotTrade order so operators can see whether the ensemble added information or just diluted conviction.
A deterministic risk layer validates the model's proposed orders against cash, leverage, symbol, concentration, and loss limits before anything reaches execution. The language model generates intent, but it cannot waive portfolio rules. BotTrade is a direct implementation target because scenarios expose constraints and the simulator records rejection, fill, equity, liquidation, and drawdown evidence for the governed policy.
Multiple prompt policies are treated as candidate programs, evaluated on completed runs, and retained or revised based on explicit objectives. DSPy provides a general framework for optimizing language-model programs, but financial use requires strict train and evaluation separation. Feed the selected policy into BotTrade only after defining the metric set, then inspect trades and risk rather than optimizing exclusively for return.
Start with the smallest architecture that can produce a valid decision, preserve portfolio state, and explain its order. Add specialists, memory, or critics only when run evidence identifies a specific failure. BotTrade's MCP and SDK interfaces support both compact and multi-agent systems while preserving the decision, trade, equity, and risk record needed to judge whether the added machinery helped.