Technology RadarTechnology Radar

Google Antigravity

ideagentmulti-agent
Assess

Google Antigravity is Google's agent-first IDE, launched in Public Preview in January 2026 alongside Gemini 3. It reframes the developer's role from code writer to "mission controller" — you issue directives, the IDE's agents plan, execute, and verify across files, terminal, and browser simultaneously.

Why It's in Assess

Antigravity is compelling in three ways: Google scale and backing, genuine agent parallelism, and free pricing. Its SWE-bench Verified score of 76.2% places it among the highest-performing publicly benchmarked coding agents — well ahead of where Devin debuted (13.86% in 2024) and competitive with Cursor and Codex. It is free for individual developers and supports multiple models (Gemini 3.1 Pro/Flash, Claude Opus/Sonnet 4.6, GPT-OSS-120B), making it a low-friction trial.

Assess rather than Trial because:

  • Still in Public Preview (not GA) as of March 2026
  • Released only in January 2026 — limited production case studies
  • Google's previous developer IDE efforts have a mixed track record (Cloud Shell Editor, Project IDX → now a competitor to this)
  • The "agent-first" paradigm it embodies is genuinely new; teams need evaluation time to understand the workflow shift

Watch for a GA announcement and independent benchmarks on real-world tasks before moving to Trial.

Two Distinct Interfaces

Editor View — a familiar AI-powered code editor with tab completions and inline commands for synchronous, file-by-file work. This is the on-ramp for existing VS Code / Cursor users.

Agent Manager (Manager View) — a task-oriented interface where you issue high-level directives and the IDE dispatches one or more agents to execute them. Agents generate Artifacts (task plans, implementation summaries, browser recordings) so you can verify the agent's reasoning without reviewing every diff. You can comment on an Artifact like a document and the agent incorporates feedback without halting.

Multi-Agent Parallelism

The Manager View supports dispatching multiple agents simultaneously — e.g., five agents on five different bugs — each with its own context and execution thread. This is the primary differentiator over single-agent tools like Claude Code CLI or standard Cursor.

Security Model

Antigravity enforces safety via a Terminal Command Auto Execution policy chosen at setup, plus per-project Allow/Deny lists for shell commands. This is application-layer enforcement, similar in approach to Claude Code's permission model (rather than OS kernel-level sandboxing like Codex CLI).

Key Characteristics

Property Value
Release January 2026 (Public Preview v1.20.5 as of March 2026)
Developer Google
Primary model Gemini 3.1 Pro / Gemini 3 Flash
Also supports Claude Opus 4.6, Claude Sonnet 4.6, GPT-OSS-120B
SWE-bench Verified 76.2%
Pricing Free for individuals (Public Preview); enterprise pricing unannounced
Platforms macOS, Windows, Linux
Website antigravity-ide.com

Further Reading