You ask your AI coding assistant to "add authentication to the backend and write tests for it." What you get is a polite suggestion to break it down into smaller prompts, or worse, a hallucinated npm install command that does precisely nothing. The common fix? You sigh, open three more tabs, and start manually orchestrating the steps yourself—defeating the entire point of having an AI helper in the first place.
Google's Gemini CLI has been listening. With releases v0.31.0 and the more substantial v0.32.0 (from March 3, 2026), it's not just offering a smarter parrot for your terminal. It's deploying a cadre of specialized agents designed to handle ambiguity, take direction, and execute complex, multi-faceted tasks. This shifts the tool from being a reactive command-line interface to a proactive development partner.
The Generalist: Your New Terminal Traffic Cop
The headline act in v0.32.0 is the Generalist Agent. Think of it as the project manager living inside your ~/.gemini directory. Instead of you having to decompose "scrape the latest Playwright release notes and summarize the testing changes" into discrete curl, jq, and analysis steps, the Generalist Agent handles the routing and delegation.
It intelligently decides what needs to be done and calls upon the right capabilities. This is enabled by default, but you can see its handiwork in a more sophisticated Plan Mode. Let's ask it to do something useful for a QA engineer.
gemini plan "Check if the login endpoint on https://demo.playwright.dev is up, and if it is, draft a basic Playwright test for a successful login."
In versions prior to v0.32.0, this might have resulted in a generic script or a request for clarification. Now, the CLI will generate a multi-step plan, which you can open in your external editor (a new enhancement) to review and tweak before execution. The plan might look like:
steps:
- agent: browser
action: navigate_to https://demo.playwright.dev/login
goal: Verify page loads and login form is present
- agent: generalist
action: analyze_page_structure
goal: Identify selectors for username, password, and submit button
- agent: code
action: generate_playwright_test
goal: Create a test file with the extracted selectors and assertion
You can then execute this plan with gemini run, and the agents work in concert. No more context switching between a browser, your IDE, and the terminal.
Taking the Wheel: Model Steering and the Browser Agent
If the Generalist Agent is the conductor, Model Steering is how you tune the orchestra. Also in v0.32.0, this feature lets you inject high-level directives into your workspace configuration, shaping the AI's behavior for specific projects.
Create a .gemini/workspace.yaml file:
model: "gemini-2.0-flash-exp"
steering:
focus: "security and testing"
style: "concise, code-first explanations"
constraints: "always propose unit tests for new functions"
Now, every interaction within this directory is subtly biased. Ask it to "add a user profile API," and its first response will likely include not just the route handler but also validation logic and a suggestion for integration tests. It remembers the mission.
Then there's the Experimental Browser Agent (from v0.31.0), which is a game-changer for QA. It allows the CLI to interact with real web pages. Need to verify a UI change or pull data for a test fixture? You can now do it from the command line without Selenium or Playwright scripts initially.
gemini browse "Go to https://api.github.com/repos/microsoft/playwright/releases/latest and tell me the published_at date and tag_name."
The CLI will fire up a headless browser, navigate, extract the information, and present it cleanly:
[Browser Agent] Navigating to https://api.github.com/...
[Browser Agent] Page loaded. Parsing JSON response.
Result:
- tag_name: v1.47.0
- published_at: 2026-02-15T10:30:00Z
Task completed successfully.
This isn't just simple curl-based API scraping; it can handle JavaScript-rendered content, click buttons, and fill forms. Imagine automating a quick smoke test of a staging environment directly from your CI/CD pipeline's shell script.
The Safety Net: Automated Review in Conductor
Perhaps the most significant update for production confidence is the Automated Review feature, part of the Conductor extension announced alongside v0.32.0. After Gemini CLI generates or modifies code, you can now command it to self-audit.
Using Conductor's workflow, you add a verify step. For example, after generating a new function, you'd run:
gemini conductor verify --path ./newAuthModule.js --guidelines "no console.log in production code, all functions must have JSDoc, use async/await error handling"
The tool then produces a report, highlighting violations, potential bugs, and adherence to your specified rules. It's like having a relentless, instant code reviewer that never complains about your variable naming (though it might silently judge your fooBarBaz choices). This addresses the fundamental trust issue with AI-generated code: it moves verification from a hopeful manual check to a systematic, automated gate.
The Payoff: Less Thinking, More Doing
The cumulative effect of these updates is a terminal that absorbs complexity. Parallel extension loading in v0.32.0 means the CLI itself starts faster, removing a minor but annoying friction point. The Plan Mode enhancements mean you're in control of complex tasks without micromanaging them.
For the developer, it means drafting a feature and its tests can be a single conversation. For the QA engineer, it means environmental checks, data gathering, and even test generation can be initiated with a natural language command, bypassing the boilerplate. The CLI is evolving from a tool that answers questions to one that accomplishes objectives.
The old way—manual decomposition, script writing, and constant vigilance—is officially tedious. Gemini CLI v0.32.0 and its agentic cohort offer a path where the machine handles the logistics, and you steer the strategy. Update your installation, define a workspace, and give your terminal a promotion from assistant to agent. Just don't be surprised when it starts asking for a corner office.