2025, the year of autonomous coding agents

If last year in AI could be called the year of chat assistants and RAG, 2025 can confidently be called the year of autonomous AI agents. In software development, the most noise at the start came from Devin, an autonomous agent focused on solving a task completely, from the task description to a PR in your repository. For each task it spins up an isolated environment that brings together the now-familiar chat, IDE-based development tools, your instructions and, of course, varied access to company knowledge through RAG. The developer can step in at any stage of the task if needed.

But at the core of such an agent are general-purpose models like Claude Sonnet 3.7, armed with various instructions and access to tools, and working on top of reasoning. So the implementation is fairly easy to reproduce, and in 2025 similar solutions started pouring out as if from a horn of plenty. For example, JetBrains released Junie, an agent integrated into its IDEs, and Cursor released a dedicated agent mode.

And last month the big players joined the fight for the lead in this area. OpenAI released its cloud version of an autonomous agent, Codex, and Google answered almost immediately with its own solution, Jules. This is the next step in the evolution of autonomous agents. Codex runs not on general-purpose models but on a specialised model for development tasks, fine-tuned through reinforcement learning. The first hands-on trials showed that results on the same tasks differ quite a lot. The UX is still very rough, but the overall logic is based on autonomous, isolated environments for each task with all the necessary tools inside. Many tasks can run in parallel, and solving one task can spawn a batch of new ones. The goal is full autonomy, where the agent always aims to create a PR that solves the task, not just a set of recommendations or changes for the developer.

I think OpenAI will spend a couple of months polishing the customer experience and shaping a commercial offer. For now you can try it for free within a ChatGPT business account (Team or Enterprise). For those who want to work locally, there’s only a less powerful and less visually convenient option so far, Codex CLI.

Software development seems to have become the first big arena for autonomous agents. Which clearly doesn’t improve the situation on the job market and with salaries in the IT sector…