Modern financial markets demand systems capable of coordinated decision-making across multiple domains. A multi-agent trading framework is presented in which five LLM-powered specialists—Portfolio Selector, Quantitative Analyst, Strategy Architect, Execution Strategist, and Risk Architect—are orchestrated sequentially from stock selection through risk assessment. Coordination is effected via CrewAI, and guardrails are applied through schema-constrained outputs and ticker-validation routines to curb hallucinations and enforce determinism. Comprehensive experiments with GPT-3.5-turbo span temperatures (0.3–0.7), risk profiles, and market regimes using 2024 data. Backtests report a 21.54% average annual return with Sharpe \(>1.0\) , while sector diversification is maintained and recommendations adapt to risk preferences. A clear temperature–behavior trade-off is observed: at 0.3, stable large-cap portfolios are produced (78% Jaccard similarity across runs); at 0.7, more exploratory growth exposures are discovered with higher volatility. The pipeline operates at approximately $0.05 per recommendation, enabling cost-aware deployment. Transaction costs and stress-test considerations are incorporated as deployment caveats and discussed alongside practical guardrails. An open-source implementation is released to demonstrate that multi-agent LLM architectures can deliver sophisticated, auditable strategies previously associated with professional trading teams, establishing a foundation for collaborative AI in financial markets.