When Google published its March 10 blog post on Gemini Code Assist, the announcement covered two distinct additions to the IDE extensions: Auto Approve for Agent Mode and two new features called Finish Changes and Outlines. That's the surface read. The more honest read: Google is admitting the code generation problem is solved. The review problem isn't.
Auto Approve — added March 10 — is Google's bet that reducing review friction is where the real gains are. The feature lets the agent propose a plan, get one approval, and execute across multiple files without stopping at each change. Disable it and you're back to plan-first by default. Agent Mode, which powers the agentic execution layer, was announced July 17, 2025, with earlier access available to Insiders in June before broader preview availability, per Google's own posts.
"Auto Approve, inline diffs, and the Context Drawer target the comprehension gap that slows AI-assisted development: not generating changes fast enough, but reviewing, validating, and trusting them," DevOps.com noted. That's a meaningful shift in how Google is positioning its AI coding tool — and it reflects something real in the IDE market.
The second addition is Finish Changes and Outlines, both now generally available in the IntelliJ and VS Code extensions, powered by Gemini 3.0. Finish Changes hit GA for IntelliJ on February 24 and VS Code on March 4; Outlines followed the same path, according to Google Cloud's release notes. Preview started December 5. The March 10 post on the Google Developers Blog is a formal writeup of shipped code — not a launch announcement. The team: Divyansh Chaturvedi (senior product manager), Nikhil Kapoor (engineering manager), and Kensen Shi (research engineer), per Google's own post.
Finish Changes works differently than Copilot or Cursor. Those tools start from prompts: you describe what you want, the model writes it. Finish Changes starts from code you're already writing. Make a partial edit, drop a // TODO:, write a few lines of pseudocode — Gemini infers the rest. Four workflows, no prompt required beyond the initial gesture: implementing pseudocode, applying patterns, following // TODO: instructions, and completing or refactoring partial code, as Google's documentation describes it.
The propagation is the point. Google's own documentation says Finish Changes applies that pattern everywhere else it is needed in the file — within-file reference propagation, with other open files used as context. The interaction model differs from prompt-first tools regardless: one manual change, Gemini fills in the rest within scope. Whether that scope holds on a large, messy codebase with non-obvious patterns is the open question.
Outlines generates concise English summaries interleaved directly in source code — inline, aligned to the code it describes. Interactive navigation is part of it: click an outline item, the editor scrolls to the corresponding block. The summary regenerates when the code changes. Both features are available to all Gemini Code Assist users through the VS Code and IntelliJ extensions with no additional subscription tier.
The competitive context: GitHub Copilot has 20 million all-time users and 90 percent Fortune 100 adoption, as Microsoft CEO Satya Nadella told TechCrunch in July 2025. Cursor has reportedly surpassed $2 billion in annualized revenue, per Bloomberg, reported by TechCrunch in March 2026. Google Code Assist went free for individual developers in February 2025 with 180,000 code completions per month, a tier it markets as 90x more than GitHub Copilot's free option, per Google's own announcement. The pricing play was aggressive. The features are catching up.
Finish Changes plus Auto Approve is a coherent IDE strategy, not two unrelated feature drops. One reduces the prompt burden by working from code already written; the other reduces the review burden by executing a plan across files with a single approval. Google stopped competing on generation speed and started competing on the full development workflow. Whether Gemini 3.0's inference quality holds up on real codebases — large ones, messy ones, with non-obvious patterns — is the question both features share. Google has the workflow in place. The inference is what they'll be judged on.
