2025 marked significant developments in the world of large language models (LLMs), focusing on key trends like reasoning, agents, and coding capabilities. OpenAI’s introduction of reasoning models, such as o1 and o3, set the stage for a revolution in how LLMs approach problem-solving. These models leverage Reinforcement Learning from Verifiable Rewards (RLVR), enabling them to tackle complex tasks by breaking problems into manageable steps. The result has been a marked improvement in areas like AI-assisted search and debugging code, allowing models to effectively diagnose issues within large codebases. Agents emerged as a notable trend this year, defined as LLMs that execute multi-step tasks using tools. While initially met with skepticism, particularly regarding their reliability, by the end of the year, agents demonstrated significant utility, especially in coding and information gathering. The introduction of Claude Code in February was a game-changer, representing a new class of coding agents capable of writing and executing code autonomously. This year also saw the rise of several vendor-independent coding agents, like GitHub Copilot CLI and OpenHands CLI, which integrated into popular development environments. The article emphasizes the shift towards asynchronous coding agents, allowing users to submit tasks and receive results later, enhancing productivity and addressing security concerns associated with running code on personal devices. This trend reflects a broader move towards more efficient, user-friendly tools in AI development, illustrating how LLMs are reshaping the landscape of programming and research. The advancements made in 2025 highlight a growing capability for LLMs to not only assist but also autonomously perform tasks that previously required human intervention.