A Comparative Evaluation of CrewAI, LangGraph, OpenAI Agents SDK, and Smolagents for Autonomous Stock Portfolio Analysis on the Helsinki Stock Exchange
Year: 2026
Type: dissertation
Abstract: In this study, we designed, implemented and evaluated four autonomous, multi-agent AI systems to analyse stock portfolios on the Helsinki Stock Ex-change. As frameworks based on large language models become more popular, professionals in the field are finding it difficult to determine which framework is best suited for domain-specific financial analysis tasks. To find out which one perform better for this purpose we selected four frameworks. Each of them represents a different architecture for comparison: CrewAI, LangGraph, the OpenAI Agents SDK and Smolagents. Each system used the same data sources, financial analysis tools and GPT-4o language model to analyse ten OMXH-listed stocks over six months of historical price data. We evaluated the performance of the systems across three categories: System performance metrics, like latency and token consumption; Structural output quality metrics, measuring report completeness and financial metric coverage; and A lightweight backtest, here we did compare each framework's recommendations against real OMXH market returns over a three month period. After we ran the apllication we built, the results show that no single framework is superior in all areas. For example, LangGraph was the fastest, taking 11.19 seconds. CrewAI produced the most detailed reports, achieving a perfect structural quality score. In the other hand the OpenAI Agents SDK achieved the highest backtest score of +56.94%, which means the strongest alignment with real market outcomes. Smolagents performed poorly across all layers.
Source: Theseus (Ammattikorkeakoulujen)
Author: Alatas, Davuthan
Institution: Satakunta University of Applied Sciences
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7162972547