Code Large language models (Code-LLMs) have made significant strides in supporting code intelligence tasks such as code generation and program repair. Despite these advances, a substantial gap remains between the capabilities of Code-LLMs and the expertise of top-tier software engineers. Agentic AI, which enables iterative reasoning and communication among agents, offers a promising avenue by allowing multiple rounds of generation and refinement. However, the foundational Code-LLMs still fall short in terms of reasoning and engineering acumen.

In this talk, I will discuss the current limitations of Code-LLMs (even when used in an agentic framework), with a focus on gaps in reasoning and communication. I will then present our recent work on clarifying and enhancing these aspects within both the agentic framework and the underlying models. The talk will conclude with insights and future directions for strengthening the agentic pipeline to better support code generation tasks.