Google is giving its Gemini CLI a much more thoughtful “measure twice, code once” brain with the rollout of plan mode, a new read-only mode designed to help developers research, reason, and design before a single line of code is touched. Instead of jumping straight into auto-editing your files, Gemini now has a dedicated planning phase where it can explore your codebase, map dependencies, and ask clarifying questions, all without the risk of it silently refactoring something you didn’t mean to change.

At its core, plan mode is a safety rail and a workflow upgrade rolled into one. It restricts Gemini CLI to a limited toolset: it can read files, run grep-style searches, glob through directories, and consult documentation, but it can’t modify anything except its own internal plan artifacts. That creates a clear separation between “thinking” and “doing” — you get a detailed strategy first, then explicitly opt in when you’re ready for the agent to edit code. For teams wary of letting an AI loose on a large monorepo, this is an important trust-building step.

Where this gets interesting is how interactive the planning flow is. Google has paired plan mode with a new ask_user tool, which effectively formalizes what should have been happening in AI coding sessions all along: the agent stops, asks you targeted questions, and only proceeds when requirements are clear. Instead of making up architecture decisions, Gemini CLI can ask whether you prefer a migration approach A vs B, which database to target, or where the real configuration file lives in a legacy project. The end result is a plan that feels more like a collaborative design doc than a black-box suggestion.

Plan mode also isn’t confined to your local filesystem. Thanks to support for read-only MCP tools, Gemini CLI can pull in context from the rest of your developer stack — GitHub issues, Postgres schemas, Google Docs, and more — but still remain in a non-destructive state. That makes it particularly useful for bigger refactors or cross-service feature work, where understanding how systems talk to each other is more important than writing the first function. This is the kind of scenario where traditional “chat with your code” setups often fall down because they lack structured, multi-source context.

On the ergonomics side, Google is making plan mode feel like a first-class citizen rather than an obscure flag. It’s enabled by default for all users, and there are multiple ways to invoke it: type /plan in the Gemini CLI input, press Shift+Tab to cycle approval modes until you land on Plan, or just ask Gemini to “start a plan for…” a task you have in mind. If you live in the terminal and want to build planning into muscle memory, you can flip a setting so new sessions start in Plan by default via /settings and choose Plan as the Default Approval Mode.

Under the hood, plan mode also ties into Gemini’s model-routing story. During planning, Gemini CLI automatically routes to higher-reasoning Pro-tier models, including the newer Gemini 3.1 Pro, to handle the heavier architectural thinking. Once a plan is approved and you move into implementation, the CLI can switch to a faster Flash-style model to actually execute edits at speed, balancing quality and responsiveness without forcing you to think about which model to pick.

If you want more than a single session plan, Google is positioning plan mode as a foundation for richer workflows, not just a toggle. A good example is Conductor, a Gemini CLI extension that already leans heavily on planning concepts like “tracks,” specs, and Markdown-based plans. Conductor effectively turns AI-driven development into a structured pipeline: define context, generate specs and plan files, then implement step by step while keeping the human developer in charge. With the first-class plan mode and the ask_user tool, Conductor can now run exhaustive pre-flight checks, confirm decisions at milestones, and then hand off to editing modes with less risk.

Extensibility is a recurring theme. Google’s documentation calls out that you can layer your own policies and workflows on top of plan mode, using tools like enter_plan_mode and exit_plan_mode and custom policies to enforce team-specific rules. For example, you could require a plan step for touching certain critical services, or wire plan mode into a homegrown extension that enforces test-driven workflows. With Gemini CLI extensions installable via a simple gemini extensions install <URL> command, the ecosystem around these planning-focused workflows should be straightforward for teams to experiment with.

Importantly, Google is also acknowledging that not everyone wants another mode in their stack. If you prefer a lean terminal or already have a battle-tested planning process, you can disable plan mode entirely from /settings by searching for “Plan.” Doing so removes it from the Shift+Tab rotation and unregisters the plan-related tools, leaving you with the existing Default and Auto-Edit (or “YOLO”) style flows. The message is clear: Gemini CLI is meant to adapt to your workflow, not force a particular process on you.

From a broader perspective, plan mode nudges AI-assisted development closer to how experienced teams already ship features: gather context, discuss tradeoffs, write a plan, then implement. Traditionally, most AI coding tools have optimized for “just write the code,” which is great for small tasks but dangerous for multi-service changes, migrations, or anything that requires real systems thinking. By formalizing planning as a dedicated, read-only phase with high-reasoning models, integrated Q&A, and extension hooks, Google is betting that the next wave of AI coding productivity comes from better process, not just faster code generation.

If you’re already using Gemini CLI, this release doesn’t require any special setup: plan mode is live and ready to try on your next feature, refactor, or migration. The more interesting question is how quickly teams will bake it into their standard dev playbooks — for example, “all risky changes must go through a plan” — and how the extension ecosystem evolves around it. For now, it’s a strong signal that AI tooling is maturing beyond autocomplete: from just writing code to actually helping you decide what to build and how to build it.