Google Antigravity CLI 2.0: coordinating coding agents from the terminal

Suppose a repository needs a framework migration, a round of tests, and a security review. Giving one assistant the whole pile can feel like hiring one person to cook dinner, fix the plumbing, and explain the smoke alarm at the same time. Google's Antigravity CLI points toward a different setup: use a terminal-based agent platform to coordinate specialized agents, while keeping a developer responsible for the work.

Google announced Antigravity 2.0 and a new Antigravity CLI at its May 2026 developer keynote. The company says the tools can launch specialized subagents for complex workflows. It also describes cross-platform terminal sandboxing, credential masking, and hardened Git policies as built-in protections. Those details make the CLI worth a look for teams that already work in terminals, but they do not answer every practical question about cost, data handling, or how much supervision a real project needs.

Read Google's official Antigravity 2.0 announcement for the launch context and the other developer tools introduced at I/O.

What Antigravity CLI is meant to do

Antigravity is Google's agent-first development platform. The CLI is a terminal surface for working with it, alongside the platform's other interface. In broad terms, the idea is to give an AI agent a coding task, let it work through a repository, and split some larger jobs among more specialized agents. Google describes the system as a way to orchestrate and build agents, rather than a code-completion feature alone.

That distinction matters. A completion tool suggests a line or function while you stay in the driver's seat. A coding agent can take a task, inspect files, make changes, and run steps in a workflow. Delegation can save context switching, but it also means the developer has to inspect the plan and review the result. A faster way to produce a patch is still a way to produce a patch that needs review.

Where a terminal coordinator could help

Imagine a team preparing to move a web app from one framework to another. One agent might inspect the routing and list likely migration work. Another could focus on tests or documentation. A developer can use those separate findings to plan a sequence of smaller changes instead of asking one agent to rewrite the project in a single leap.

That is an example of how subagent orchestration could be used, not a claim that Google's launch announcement demonstrated this exact migration workflow. Google's keynote also mentioned an automated Migration agent for converting frameworks to native Kotlin, a separate Android-focused announcement. Teams should keep these features distinct when deciding what the Antigravity CLI itself can do.

Smaller tasks are usually easier to review. A good first assignment might be to map a package's callers, add a focused test, or identify files touched by an upgrade. Give the agent a clear boundary, ask for a summary of changes, and check the diff before merging. For a task with multiple independent pieces, parallel agents may help, but only if their edits do not collide and someone owns the final review.

Why the safety details deserve attention

Terminal access is useful because development work lives in files, commands, and version control. It is also where mistakes can have consequences. A command can overwrite work, install dependencies, or send information to a service. Google says Antigravity includes terminal sandboxing across platforms, credential masking, and hardened Git policies. Those are sensible controls to investigate, not a reason to assume every repository or command is safe by default.

Before using an agent on a real project, ask what the sandbox allows it to read and change, which network access is available, and how secrets are handled. Keep credentials out of prompts and test with a low-risk repository first. Use a branch or other recoverable copy, review the proposed commands, and inspect the Git diff. If the product exposes permission settings, learn how to limit file and command access before granting broader access. Google's keynote summary does not spell out every setting or policy behavior, so verify those details in the current product documentation.

There is a useful human rule here: an agent should not get more access than its assignment needs. A task to summarize a module rarely needs permission to modify the whole repository. A request to edit code does not automatically need access to deployment credentials. Sandboxing helps, but a narrow task and a careful operator are still part of the safety plan.

What the announcement does not settle

The keynote is a launch overview, not a complete operating manual. It does not give a price for Antigravity CLI, explain usage limits, or describe the full data-retention and model-routing policies. It also does not establish whether every capability is available to every account or platform. Check the current Antigravity terms and documentation before making a budget or privacy decision; do not infer those answers from the word sandbox.

For private or regulated code, a team should confirm where prompts and repository context go, what gets retained, and whether its plan offers the controls it requires. For cost, check whether usage is included in an existing plan or billed separately, and test a representative workflow before estimating monthly spend. These questions are ordinary procurement homework, not an accusation that the product handles data badly.

How it compares with other coding assistants

Antigravity CLI belongs in the same evaluation conversation as terminal coding agents and IDE assistants, including Gemini CLI, Claude Code, and GitHub Copilot. Those products do not have identical workflows, permissions, or pricing, so a feature checklist alone can mislead. A tool that suits a terminal-heavy team may feel clumsy to a group that works mainly in an IDE. Likewise, a multi-agent workflow can be more machinery than a small project needs.

Use one real task to compare them. Pick a bounded issue with a known expected result, such as adding a test for a bug or updating a small integration. Note how much setup each assistant needs, whether it explains its changes, how easy it is to restrict access, and how long review takes. Include the time spent correcting the output. If a tool writes code quickly but leaves a confusing diff, the apparent speed may disappear during review.

Who should try it first

Antigravity CLI is most interesting to developers who already live in a terminal and want to explore delegated coding tasks or agent orchestration. It may also suit teams that are deliberately testing how agent workflows fit into their engineering process. Readers looking for a basic autocomplete helper, a fully local model, or a clearly documented fixed price should verify those requirements before adopting it. Google's announcement alone does not establish those capabilities.

Start with a disposable or low-risk repository, assign one narrow task, and keep the change behind a branch. Then judge the result by the work you had to review, not by how impressive the agent's activity looked in the terminal. If it produces a clean, understandable patch and stays inside the permissions you gave it, try a slightly larger task. If you cannot explain what it changed or why, pause and tighten the workflow before handing it more of the codebase.