Google debuts self-coding AI Jitro

Coding agents are software tools that do more than autocomplete: they can plan a task, edit files, run tests, read errors and try again until the code works. Google has been building that kind of system into Jules, its asynchronous coding agent, and recent reporting says the next step is a project called Jitro. (cloud.google.com) (blog.google) (devops.com) Google has not published a standalone official announcement for Jitro on its main blogs or Labs pages that were available on April 12, 2026. The company’s public coding products are still branded around Jules and Gemini Code Assist, while outside reports describe Jitro as a more autonomous follow-on to Jules. (blog.google) (labs.google) (testingcatalog.com) Jules entered public beta on May 20, 2025, and Google said it could clone a repository into a secure Google Cloud virtual machine, understand the project context and handle jobs such as writing tests, fixing bugs and updating dependencies. Google later said Jules was available publicly and had produced more than 140,000 code improvements during beta. (blog.google 1) (blog.google 2) On Google’s current Labs page, Jules is sold in tiers that range from 15 tasks a day and 3 concurrent tasks to 300 tasks a day and 60 concurrent tasks, with higher plans getting earlier access to newer Gemini models. That matters because Google is already treating coding help less like a chat box and more like background work that runs in parallel. (labs.google) Google has been moving in that direction for months. In a December 2025 update, the company added “Suggested Tasks” that scan repositories for improvements and “Scheduled Tasks” that run maintenance jobs on a cadence, including dependency checks and other recurring work. (blog.google) The technical idea behind a self-improving coding system is simple: the model writes code, software checks whether that code works, and the model uses the result to make another pass. Google described a similar loop in AlphaEvolve, a Gemini-powered research agent announced on May 14, 2025, which generated programs, scored them with automated evaluators and kept the stronger candidates. (deepmind.google) Google said AlphaEvolve had already been used inside its own computing systems, including data centers, chip design and artificial intelligence training. That shows the company has been testing “write, verify, improve” systems in-house before pushing more autonomous coding tools toward developers. (deepmind.google) Google is also expanding the surrounding developer stack. Gemini Code Assist became generally available on May 20, 2025, and Google said an internal experiment found developers using it were 2.5 times more likely to complete common development tasks successfully than developers without coding assistance tools. (blog.google) At the same time, Google’s own 2025 DORA survey of nearly 5,000 technology professionals found 90% were using artificial intelligence in software development, more than 80% said it improved productivity and 30% said they trusted it only “a little” or “not at all.” That gap helps explain why Google’s public messaging around Jules keeps emphasizing review, approval and isolation of private code. (blog.google 1) (blog.google 2) So the clearest reading of the Jitro story is not that Google suddenly invented self-writing software this week. It is that Google’s public coding tools, internal research agents and recent product updates all point in the same direction: less prompting, more delegation, and more code that gets written, tested and revised by the system before a developer signs off. (blog.google) (deepmind.google) (blog.google)