This paper explores the use of heterogeneous large language models (LLMs) for financial sentiment analysis (FSA), a task that has traditionally relied on fine-tuning pre-trained models. The study proposes a design framework called Heterogeneous Agent Discussion (HAD), which leverages Minsky's theory of mind and emotions to create specialized LLM agents. These agents are designed to focus on specific types of errors in FSA, such as irrealis mood, rhetoric, and aspect mismatch. The framework evaluates the effectiveness of HAD through empirical testing on five FSA datasets, ablation analysis, and case studies. Results show that HAD improves FSA accuracies, especially when the agents produce substantial discussions. The study contributes to the design science literature by providing a kernel theory-informed design artifact and offers implications for business and management practices in financial decision-making.This paper explores the use of heterogeneous large language models (LLMs) for financial sentiment analysis (FSA), a task that has traditionally relied on fine-tuning pre-trained models. The study proposes a design framework called Heterogeneous Agent Discussion (HAD), which leverages Minsky's theory of mind and emotions to create specialized LLM agents. These agents are designed to focus on specific types of errors in FSA, such as irrealis mood, rhetoric, and aspect mismatch. The framework evaluates the effectiveness of HAD through empirical testing on five FSA datasets, ablation analysis, and case studies. Results show that HAD improves FSA accuracies, especially when the agents produce substantial discussions. The study contributes to the design science literature by providing a kernel theory-informed design artifact and offers implications for business and management practices in financial decision-making.