Google Antigravity 2.0: The Harness You're Not Using (Yet)

A frontier language model on its own can only read tokens and output tokens.

To do real software engineering, it needs a harness: the runtime layer that controls how the model reads files, executes terminal commands, manages long-running tasks, respects workspace boundaries, and verifies its own output.

Antigravity 2.0 approaches this as an integrated environment across four surfaces:

• a standalone desktop application for orchestrating agent threads
• a full IDE experience for direct code editing and diff review
• a command-line interface (agy) for terminal-first workflows
• cloud sandboxes and the Interactions API in Google AI Studio

One of the most practical parts of Antigravity 2.0 is how it handles command execution and file permissions.

By default, terminal commands run inside a local sandbox scoped strictly to your active workspace directory, scratch folders, and standard build tools—without unrestricted network access or access to unrelated files on your machine.

When a command genuinely needs external network access—such as pushing to GitHub, deploying to Vercel, or calling an outside API—the harness requests explicit approval to bypass the sandbox for that specific step.

That separation lets routine file reads, lint checks, and local builds run automatically at full speed without giving an agent unchecked access to your entire operating system.

A common mistake when configuring coding agents is dumping thousands of lines of instructions into a single global prompt file. That bloats the context window on every turn and degrades reasoning quality.

Antigravity 2.0 solves this with a modular Skills architecture built around progressive disclosure.

Each skill lives in its own folder with a required SKILL.md file (containing YAML frontmatter with a concise name and trigger description) alongside optional helper scripts, examples, and reference docs.

At the start of a session, the agent only sees the short one-line descriptions of available skills. It only reads the full SKILL.md file when your task actually requires that specific capability—keeping the context window lean until specialized knowledge is needed.

Instead of forcing you to manually prompt every follow-up step, Antigravity builds structured workflows into slash commands:

• /goal turns an open-ended prompt into an autonomous execution loop where the agent plans, implements, runs builds and tests, and does not stop until the completion criteria are verified.
• /learn captures corrections, environment quirks, or design preferences from the current session and persists them as reusable rules or skills so you never have to repeat the same fix twice.
• /teamwork coordinates multiple autonomous subagents working in parallel across isolated Git branches or worktrees, allowing one agent to handle frontend components while another builds data pipelines or runs verification.

Terminal build passes do not guarantee that a web page actually looks right.

A Next.js build can exit with code 0 while a hero section overflows on mobile, an image fails to load, or a layout looks generic.

Antigravity integrates directly with Chrome DevTools via Model Context Protocol (MCP). An agent can launch a real headless or visible browser tab, navigate to your local dev server, inspect the live DOM and console errors, test responsive viewports, and capture screenshots to visually verify its own frontend changes before handing the result back to you.

Beyond interactive coding sessions, Antigravity extends into persistent background automation:

• Scheduled tasks let you run recurring cron jobs or one-shot timers that wake the agent up to check data feeds, run audits, or summarize updates.
• Mobile remote control pairs via QR code so you can monitor a running local session, review diffs, or approve a sandbox permission request from your phone.
• On the cloud side, Google AI Studio's managed sandboxes and the new stateful Interactions API make it straightforward to move agentic workflows from local experimentation into production applications.