Embedded software development in the automotive sector is becoming more demanding as engineers balance performance, safety, and certification requirements. Traditional approaches such as manual coding or model-based workflows require significant effort in writing boilerplate code, ensuring AUTOSAR compliance, and preparing verification artifacts for audits. Commercial tools like MathWorks Simulink Embedded Coder already provide model-to-code generation capabilities, but these solutions are often costly and may not be viable for budget-sensitive domains such as two-wheelers and three-wheelers. This creates a strong need for alternative, more accessible approaches that can deliver efficiency without compromising on quality.
This paper introduces an embedded coding agent designed to address this gap. The agent accepts natural language or structured prompts and generates AUTOSAR-compatible ARXML as well as embedded C and C++ implementations. Unlike general-purpose coding assistants, it is purpose-built for automotive software, tightly integrated with AUTOSAR schemas, vendor documentation, and coding standards. Its architecture combines retrieval-augmented generation, template-guided synthesis, and a verification pipeline that automatically produces MISRA compliance reports, static analysis outputs, traceability matrices, and results from software-in-the-loop and hardware-in-the-loop testing.
We benchmarked the system on driver generation, communication stack integration, and AUTOSAR component creation. Results show that the agent reliably produces compiling code, accelerates repetitive development tasks, and supports traceability for certification workflows. While expert oversight remains essential for optimization and safety validation, the embedded coding agent offers a cost-effective complement or alternative to traditional model-based tools, with particular relevance for emerging markets and two- and three-wheeler applications.