Google adds read-only Plan mode to Gemini CLI for safer codebase analysis and design work

Google separates planning from execution in Gemini CLI

Google said on March 11, 2026 that Plan mode is now available in Gemini CLI and enabled by default for all users. The new mode keeps the agent in a read-only state while it analyzes a request, inspects a codebase, maps dependencies, and proposes a strategy. The goal is straightforward: make research and architectural planning easier without letting the tool make accidental edits or trigger unwanted execution too early.

In Plan mode, Gemini CLI is limited to a smaller tool set focused on exploration. Google says the agent can navigate files, search for patterns, and read documentation, but it cannot modify project files apart from its own internal planning state. That makes the mode useful for tasks like evaluating a migration path, outlining a feature, or understanding how a complex service is wired before a human commits to implementation.

New collaboration and context features

Google paired the release with a new ask_user tool. Instead of guessing through missing requirements, the agent can now stop and ask targeted questions about architecture, hidden configuration, or product intent before finalizing a plan. That makes the workflow more collaborative and reduces the chance that a plausible-looking plan is built on the wrong assumptions.

The company also expanded Plan mode with read-only MCP tools. Google says Gemini CLI can safely pull in context from systems beyond the local filesystem, including sources such as GitHub, Postgres, or Google Docs, as long as those integrations stay read-only. In practice, that means a planning session can include issue context, schema details, or operational notes without giving the agent write access to the wider stack.

Why it matters for agentic development

Google frames Plan mode as a foundation for more structured agent workflows. Its Conductor extension already uses Plan mode and ask_user to orchestrate multi-step development tracks, and Google said the team intends to bring Conductor concepts into Gemini CLI more directly over time. Plan mode can be entered with /plan, by cycling approval modes with Shift+Tab, or by asking the agent to start a plan. Google also said higher-reasoning Pro models such as Gemini 3.1 Pro are used during planning.

• Plan mode is read-only and enabled by default for Gemini CLI users.
• The new ask_user tool lets the agent pause and request clarification before proposing a strategy.
• Read-only MCP support extends planning into systems such as GitHub, Postgres, and Google Docs.
• Google is using the feature as a building block for more structured agent orchestration.

Source: Google Developers Blog