Gemini CLI: Plan Mode Enabled by Default

Plan Mode Is Now the Default

With the release of Gemini CLI v0.34.0 on March 17, 2026, Plan Mode becomes the default operating mode for all users — a significant shift from the opt-in experimental feature introduced just six days earlier in v0.33.0. Rather than diving directly into file modifications, Gemini CLI now starts every session in a read-only, planning-first state. The agent researches the codebase, gathers context, and proposes a structured plan before the user ever approves a single change.

What Plan Mode Does

In Plan Mode, the agent is restricted to a curated set of read-only tools: read_file, grep_search, glob, and compatible read-only MCP integrations. It cannot modify files, run shell commands that mutate state, or make API calls with side effects. Instead, it uses these tools to build a comprehensive understanding of the codebase, then surfaces a clearly structured plan — with proposed changes, rationale, and potential risks — for the user to review and approve.

Read-Only MCP Integrations

Plan Mode now supports read-only MCP (Model Context Protocol) integrations, enabling the agent to gather richer context from connected developer tools without triggering any state-changing side effects. Supported integrations during planning include GitHub (reading issues and PRs), Postgres (querying schemas and data), Google Docs (reading documents), and any other MCP server that exposes read-only capabilities. This makes planning phases far more contextually aware without compromising safety.

The ask_user Tool

A new ask_user tool allows the agent to pause mid-research and request clarification from the developer before finalizing a plan. If the agent encounters ambiguity — for instance, whether a refactor should preserve backward compatibility — it can pause and ask rather than making an assumption. This reduces misaligned or incomplete implementation proposals and keeps the developer in the loop at the right moments.

Automatic Model Steering

One of the most technically interesting aspects of this release is automatic model steering during Plan Mode. Architectural reasoning during the planning phase is routed to Gemini 3.1 Pro, which offers stronger multi-step reasoning and can better handle complex dependency analysis, cross-file impact assessment, and architectural trade-offs. Once the plan is approved and implementation begins, the system automatically switches to Gemini 3 Flash for the faster, more cost-efficient execution phase. No configuration is required — the model switching happens transparently based on the current phase.

Exiting and Customizing Plan Mode

The change is opt-out, not forced. Developers who prefer direct execution — the previous default behavior — can revert via /settings in a single step by toggling the Plan Mode default. For teams that want to customize which tools are available during planning or adjust the model routing behavior, both settings are exposed in the configuration system.

Plan Mode can also be exited mid-session: once the agent presents its plan and the user approves, the session transitions out of the read-only constraint into full execution mode automatically.

Why This Matters

The decision to make Plan Mode the default reflects Google's philosophy that agentic AI tools should be deliberate by design. Directly executing file mutations on the first attempt introduces risks — partial changes, incorrect assumptions about intent, and hard-to-reverse side effects. A mandatory planning phase creates a forcing function: the agent must demonstrate understanding before being granted the ability to act.

The pace of adoption is also notable. Plan Mode went from experimental introduction in v0.33.0 (March 11) to on-by-default in v0.34.0 (March 17) — a six-day iteration cycle that reflects the team's confidence in the feature's stability and user reception.