Google adds Agent Mode to Gemini Code Assist for multi-file coding

Coding with AI used to feel like working with a brilliant intern who could only see one screen at a time. You'd generate a snippet for one file, copy it over, switch tabs, ask for the matching backend code, and spend the next twenty minutes manually stitching everything together.

Google is trying to fix that friction. Gemini Code Assist, the company’s AI programming assistant for both individual developers and enterprises, just picked up a major set of upgrades, headlined by Agent Mode, a new Gemini CLI, and a host of IDE workflow tweaks.

Individual developers can try the new features for free across Visual Studio Code and JetBrains IDEs.

An agent that reads the whole codebase

Google rolled out Agent Mode to Insiders last month, transforming Code Assist from a passive auto-complete bar into an active pair programmer.

Instead of just scanning whatever file you happen to have open, Agent Mode parses the entire codebase. It maps out your application architecture, dependency trees, and existing naming conventions.

Because it sees the full picture, you can now assign it multi-file tasks from a single prompt. Google highlighted scenarios like updating all API endpoints to a new authentication scheme, upgrading a core library while patching breaking changes, or spinning up a complete full-stack user settings page across multiple layers.

Handing full project control to an AI model sounds risky, so Google built in human-in-the-loop guardrails. The agent does not modify code blind. It first lays out a detailed plan summarizing every file it plans to touch and what changes it will make.

Developers can review the proposal, ask clarifying questions, suggest a different implementation, or deny specific edits. And if an accepted patch goes sideways, a checkpoint rollback feature introduced in June lets you revert every modified file back to its original state in one click.

Cleaning up the terminal and the IDE

Alongside the agent, Google shipped the Gemini CLI. It brings Gemini straight into the terminal, letting developers debug commands and ask questions without leaving the shell.

Inside the IDE, context controls got a meaningful tune-up. Code Assist now automatically respects your .gitignore rules and introduces support for a .aiexclude file, designed to keep sensitive credentials or messy legacy scripts away from the model. Developers can also highlight narrow snippets or attach terminal command logs directly into chat, bypassing the copy-paste routine.

Even the UI received quality-of-life polish. Code suggestions now live inside collapsible preview blocks, filenames mentioned in chat are clickable links that open right in the editor, and you can stop rambling generations with a single click.

Why it matters

We think this update shows Google finally targeting the real bottleneck in AI development: context and project-scale execution.

Single-line auto-complete is already a commodity. The real race is about orchestrating complex refactors across dozens of files without breaking production. Giving individual developers free access to Agent Mode inside both VS Code and JetBrains lowers the barrier to entry significantly, especially for those unwilling to switch their entire workflow over to standalone AI editors like Cursor.

Google was smart to make checkpoints and upfront review plans mandatory. Autonomous coding without an instant rollback button is a non-starter for real production codebases. By letting developers inspect the blueprint first and undo everything if it breaks, Google removes a lot of the fear factor.

Competitively, this is Google putting direct pressure on GitHub Copilot and specialized AI IDEs. By bringing multi-file agents, terminal integration, and codebase-wide awareness directly into standard development environments for free, Google is making it much harder for standalone wrappers to justify premium fees.