Google unveils Jitro self‑coding AI
Google unveiled Jitro, which it described as an AI capable of self‑coding and part of a move toward autonomous AI improvement systems. (x.com) The announcement framed Jitro as a step toward models that can iteratively improve their own code and tooling. (x.com)
A coding agent is software that can write code, run tools, inspect files, and try fixes in sequence instead of stopping at a single answer. Google is now describing a system called Jitro as a step toward agents that can improve their own code and tooling over repeated cycles. (x.com) Google has not published a Jitro product page or technical paper on its main public AI sites that was readily searchable on April 13, 2026. The public record so far is a social post describing Jitro as “self-coding” and tied to “autonomous AI improvement systems.” (x.com) That idea fits Google’s current product direction. Google’s Jules service says it can “proactively find and fix code improvements,” and Google’s new Antigravity platform says agents can plan, execute, and verify complex tasks across an editor, terminal, and browser. (google.com) (developers.googleblog.com) Google has also been building the pieces that make autonomous coding more practical. In a March 25, 2026 post, Google DeepMind said coding models have “fixed knowledge” and need live documentation, tools, and “agent skills” to keep up with changing software libraries and best practices. (developers.googleblog.com) In that same post, Google said its evaluation harness used 117 coding prompts and that newer Gemini 3 models improved sharply when paired with these skills and documentation-fetching tools. The company said the setup included tools called `activate_skill` and `fetch_url`, which let the model pull current information instead of relying only on training data. (developers.googleblog.com) Google has been pitching Gemini 3 as the model layer for this shift. In a November 18, 2025 developer post, Google said Gemini 3 Pro scored 54.2% on Terminal-Bench 2.0, a benchmark for using tools through a terminal, and said the model handles multi-file refactors, debugging sessions, and feature work across entire codebases. (blog.google) The plain-language distinction is that a chatbot suggests code, while an agent can keep working through a task with tools and checks. A “self-coding” system pushes that one step further by trying to modify the software stack that powers the agent itself, not just the user’s app. (blog.google) (developers.googleblog.com) Google’s public materials already stress verification, not blind autonomy. Antigravity says agents report progress with screenshots and recordings, and Jules is framed as a backlog assistant that proposes or applies improvements inside a developer workflow. (developers.googleblog.com) (google.com) What is still missing is the part that would let outsiders judge Jitro on its own terms: a paper, benchmark, safety documentation, or product documentation from Google. Until Google publishes that material, the clearest way to read the announcement is as a signal of where its coding-agent roadmap is heading, not yet as a fully documented public system. (x.com) (blog.google)
