Social robots are designed to interact with humans in natural and empathetic ways, employing advanced communication, emotional recognition, and adaptive mechanisms. These features enable their deployment across a range of domains, including education, healthcare, and entertainment. With the emergence of autonomous agents powered by large language models (LLMs), there is potential to further augment robot autonomy. LLM-based agents utilize deep learning to interpret complex data, reason, and generate contextually appropriate responses, supporting applications that require flexible language understanding, such as customer service and healthcare. Integrating social robots with LLM-based autonomous agents offers the opportunity to combine embodied interaction skills with enhanced reasoning and adaptability. Such integration may improve a robot’s ability to interpret dynamic social contexts, respond to diverse human behaviors, and plan actions more effectively in variable environments. This paper presents the design and implementation of an LLM-based agent architecture for multi-form social robots, with a focus on platforms such as Pepper and NAO. The proposed approach aims to enhance autonomous task execution, real-time decision-making, and interaction quality in social settings by leveraging the advanced language processing capabilities of LLMs within embodied robotic systems. Our evaluation of approximately 400 automatically generated tasks showed that while over 70% of tasks were partially completed, only 14.1% were successfully completed from end to end, highlighting the challenges of robust execution in simulation. Furthermore, in a memory intensive inventory task, the system demonstrated high efficiency, with memory being used in 24 out of 25 trials and reducing the median number of steps from 11 to 1 for the final report generation.