Google unveils self‑coding model Jitro
Google has launched Jitro, a model designed to write and improve its own code, signalling a step toward self‑optimizing AI systems. The announcement frames Jitro as a move toward more autonomous model development rather than just assisted coding. (x.com/i/status/2043082914988999151)
Google is pushing its coding agents past “write this function” prompts and toward systems that can work on broader software goals over time. (jules.google, blog.google) The closest public product today is Jules, Google’s coding agent for GitHub repositories. Google says Jules can fix bugs, add documentation, update apps, implement features, run multiple tasks at once, and handle up to 300 tasks a day on paid plans. (jules.google, jules.google) Jules already works more like an outsourced engineer than an autocomplete tool: a user connects a repository, assigns a task, and the agent works asynchronously in the background. Google described that shift when it launched Jules as an “autonomous agent” rather than a code-completion assistant. (jules.google, blog.google) That framing fits a broader change inside Google’s developer products. In December 2025, Google said Gemini 3 Pro was built for “complex, long-horizon tasks across entire codebases,” and tied those capabilities to Google Antigravity, its agentic development platform. (blog.google, deepmind.google) The technical idea is simple: instead of asking a model for one patch at a time, companies are trying to give it a destination and let it plan the route. In software terms, that means an agent can inspect a codebase, choose files, make edits, test them, and revise its own work before a human reviews the result. (blog.google, jules.google) Google has been moving in that direction for years in research. DeepMind’s AlphaDev used reinforcement learning to discover faster low-level algorithms, and RoboCat was presented as a robot agent that improved by generating more of its own training data. (deepmind.google, deepmind.google) Google has also shown that coding agents can be used to improve Google’s own systems. In 2025, DeepMind said AlphaEvolve, a Gemini-powered coding agent, found algorithmic and infrastructure improvements for data-center scheduling, hardware design, and AI model training. (deepmind.google) What is still missing is a detailed public technical paper or product page for Jitro itself. As of April 12, 2026, Google’s official sites describe Jules, Gemini coding tools, and agentic development platforms, while recent reports and social posts describe Jitro as an internal or newly surfaced next step beyond today’s prompt-and-execute coding agents. (blog.google, jules.google, testingcatalog.com, devops.com) That leaves the near-term question less about whether Google wants more autonomous coding systems and more about where it draws the boundary. Google’s public products already let agents edit code and manage parallel tasks; the next step is how much goal-setting, self-revision, and model-improvement Google is willing to automate inside its own stack. (jules.google, jules.google, blog.google)