Google’s Gemini Code Assist team shipped a set of updates on March 10 that target the part of development where most time is actually spent — the hands-on-keyboard coding loop. Agent Mode with Auto Approve, inline diff views, a Context Drawer for precise file management, custom commands, and a companion feature called Finish Changes collectively aim to reduce the friction between writing code and reviewing AI-generated changes.
These aren’t headline-grabbing agent announcements. They’re workflow improvements. And for teams already using Gemini Code Assist in VS Code or IntelliJ, they address the specific points where AI-assisted coding still creates drag.
Agent Mode With Auto Approve
Agent Mode lets you describe a complex, multi-step goal and have the agent propose a plan, execute it across multiple files, and present changes for review. The update adds Auto Approve — the agent acts on your behalf without pausing at each step. Once it finishes, you review the full set of changes and roll back anything you don’t want.
The use case: Large-scale changes, such as adding an API endpoint across multiple files or applying a refactoring pattern across a codebase. Stopping to approve every file edit slows work that the agent could handle in one pass. Auto Approve removes that interruption.
The caution is equally clear. The agent has access to your filesystem and terminal. Google’s documentation is explicit: Be careful about when and where you enable it. The standard approval mode — plan first, approve each step — remains the default.
Inline Diff Views
Gemini Code Assist now highlights changes directly in the editor — green for additions, red for deletions — rather than requiring a separate diff window. Inline diffs also appear in chat for at-a-glance review.
The review friction in AI-assisted coding often isn’t about the quality of the suggestion — it’s about the effort required to understand what changed. Inline diffs keep visual feedback where you’re already working. You can edit changes directly in the integrated diff view, maintaining flow without bouncing between panels.
Context Drawer
The new Context Drawer gives developers precise, temporary control over what Gemini sees. Add or remove specific files and folders from context for your current session — reducing noise, improving relevance, and avoiding context rot caused by irrelevant code that clutters the agent’s attention.
Custom Commands
Define custom slash commands or save complex prompts for routine actions — generate test files with your preferred boilerplate, refactor code according to your style guide, run pre-commit checks. Custom commands encode team-specific patterns into reusable, invocable actions, the same principle behind VS Code agent skills and plugins applied at the command level within Gemini Code Assist.
Finish Changes and Outlines
Shipped the same day in a companion announcement, Finish Changes and Outlines address two different friction points.
Finish Changes acts as an AI pair programmer, observing your current edits and completing them. Start a refactoring pattern, implement part of a function, or write pseudocode, and Finish Changes infers your intent and completes the work. It’s not autocomplete. It’s pattern recognition applied to in-progress edits across the context of what you’re doing.
Outlines generate interactive, high-level English summaries interleaved directly within source code. For complex files, Outlines provides a readable map of what the code does without requiring you to read every line. Both features are available in VS Code and IntelliJ extensions for Gemini Code Assist.
According to Mitch Ashley, VP and practice lead for software lifecycle engineering at The Futurum Group, “Google’s Gemini Code Assist updates shift focus from what agents can do to how developers work with what agents produce. 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. The Auto Approve model is worth watching closely. Trusting agents to execute across files and reviewing after the fact trades interruption overhead for deferred oversight. That tradeoff works for well-scoped tasks with recoverable failure modes.”
“Custom commands extend the same pattern emerging across VS Code plugins, Azure Skills, and Google ADK — encode team-specific knowledge into reusable actions that execute consistently rather than relying on agents to infer it.”
Why This Matters for DevOps
The AI coding market has spent the past year focused on what agents can do. These Gemini Code Assist updates focus on how developers interact with agent output — a different problem and arguably more important for daily productivity.
Auto Approve addresses permission fatigue. We’ve covered this across multiple articles — IronCurtain putting capabilities outside the model’s reach, VS Code hooks for lifecycle enforcement, Copilot’s self-review before opening PRs. Gemini takes the opposite approach: trust the agent to execute, and review after. Both models have their place depending on risk tolerance and task scope.
Inline diffs and the Context Drawer address the comprehension gap. As agents generate more changes, the ability to quickly understand what changed matters more than the speed of generation. These features are for efficiently reviewing AI-generated code, not for producing it faster.
Custom commands connect to the broader skills and plugin pattern — VS Code agent plugins, Azure Skills, and Google ADK integrations. The direction is the same everywhere: Encode team-specific knowledge into reusable actions that AI tools execute consistently.
Gemini Code Assist is available for VS Code and IntelliJ. Agent Mode, inline diffs, and custom commands are generally available.