❯ Guillaume Laforge

Getting Started with Google Antigravity

antigravity ai-agents gemini
Getting Started with Google Antigravity
Getting Started with Google Antigravity presentation at Dev With AI 2026

Earlier this week, I spoke at the first edition of the Dev With AI conference in Paris, held at the Forum des Images. The event brought together around 500 developers to exchange notes on practical AI tooling, real-world development workflows, and coding agents.

My presentation, “Antigravity, le dev agentique façon Google DeepMind”, was a fast-paced 20-minute tour of Google’s agentic development environment. I walked through how the platform is structured, from developer surfaces down to autonomous execution and safety guardrails.

Here is a breakdown of the core building blocks I presented.

1. Developer Surfaces

Antigravity gives you multiple entry points depending on how you want to work:

  • Antigravity 2.0 (Agent Manager): A desktop application acting as an agent mission control. You assign high-level goals, watch autonomous agents execute tasks in parallel, review diffs, inspect artifacts, and collaborate directly with them.
  • Antigravity IDE: An agent-centric editor built on a VS Code base, combining full editing capabilities with native agent orchestration panes and interactive sidecars.
  • Antigravity CLI (agy): A command-line tool for developers who prefer running tasks and autonomous workflows directly from the terminal or shell scripts.
  • Antigravity SDKs: Programmatic libraries (available for Python, with an unofficial SDK for Java that I wrote) that wrap the native Go execution harness so you can build custom agent automations in your own code.

2. Autonomy and Extensibility

To tackle non-trivial tasks without constant hand-holding, Antigravity combines several autonomy primitives:

  • Subagents: Specialized workers that the main agent can spin up concurrently to explore code, research documentation, or run tests in isolation without cluttering the primary context window.
  • Custom Agents: Pre-configured specialist personas defined with their own roles, system prompts, and toolsets.
  • Agent Skills: Folders containing a SKILL.md instruction file and domain resources. Skills follow a progressive disclosure model: the agent only loads detailed documentation into context when relevant to the task at hand.
  • MCP Servers: Native support for the Model Context Protocol, enabling agents to tap into external tools and datasources seamlessly.
  • Slash Commands: Quick shortcuts in chat to trigger repeatable workflows or specialized capabilities like /plan, /browser, /goal, or /generative-ui (maybe my favorite!)

3. Orchestration, Control, and Safety

Autonomy needs guardrails. Antigravity provides structured control mechanisms to keep agents safe and accountable:

  • Artifacts: Structured markdown deliverables (like implementation plans, diffs, or reports) that give you clear checkpoints to inspect and approve before execution continues.
  • Remote Control & Steering: The ability to send messages, pause, redirect, or course-correct running agents in real time (even from your phone, when you’re away from your keyboard).
  • Rules and System Directives: Persistent guidelines (like coding style guides or safety rules) that the agent must strictly follow across sessions.
  • Sidecars and UI Extensions: Interactive visual panes embedded directly into the workspace (such as browser viewers or visual canvases).
  • Execution Policies & Sandboxing: Intercepting commands and tool calls with configurable allow/deny policies to prevent unintended actions.

Slides

The presentation slides are embedded below and available on Speaker Deck:

Resources