Antigravity 2.0: Google's Agent-First Pivot in Coding

  Google I/O 2026 brought us Antigravity 2.0. It ditches the traditional IDE entirely and goes all-in on an Agent-first
  design. Here's a look at what changed, across architecture, engineering, business model, and risks.

  Architecture: From Q&A to Multi-Agent Teams

  The biggest leap is parallel multi-agent work. Instead of one-step-at-a-time, Antigravity 2.0 can now spin up
  sub-agents for frontend, backend, testing, and docs — all running at once. In one demo, 93 sub-agents fired off over
  15,000 model requests and finished a custom OS build in 12 hours. That's months of work, compressed.

  On the surface, you get three ways in. The desktop app handles orchestration and review. The rebuilt CLI shares the
  same Agent Core for a lighter touch. And the SDK lets teams wire agent capabilities into their own stack.

  Engineering: Always-On Agents and Cloud Sandboxes

  Two things stand out on the engineering side.

  Scheduled tasks. Developers can set background jobs with /schedule. Code reviews, dependency bumps, test runs — agents
  handle them unattended. These tools are starting to look less like utilities and more like long-running, schedulable
  services.

  Managed agents via Gemini API. A single API call provisions an isolated Linux sandbox with reasoning, tool use, and
  code execution baked in. Declarative configs like AGENTS.md and SKILL.md cut down the setup time. Model choice stays
  open — Gemini 3.5 Flash ships out of the box, but you can plug in Claude Opus 4.6 and others.

  Business: Selling Environments, Not Tokens

  The pricing shift is just as notable as the tech. Per-request billing is gone. It's now a compute quota system — Pro
  users get weekly limits with a five-hour refresh window. A new $100/month AI Ultra tier sits on top.

  The signal is clear. Fixed monthly subscriptions don't work when marginal costs won't trend toward zero. Google is
  moving from selling model tokens to selling agent runtime environments. The Gemini CLI was absorbed into Antigravity.
  What's emerging is something closer to Docker, but for agent operations. Users pay for orchestration, integration, and
  deliverables — not raw compute.

  Risks

  A few rough edges remain.

  The UI jump is jarring. Developers lose familiar editors, file trees, and version control. An Agent Manager replaces
  them. Without enough trust built up, the black-box scheduling has drawn real pushback.

  The new CLI went closed-source. The old Gemini CLI was open. The team promises a plugin system to absorb existing
  extensions, but smooth migration is far from guaranteed.

  Quota tracking is opaque. There's no clear feedback on compute usage. Complex tasks can quietly hit hidden caps,
  breaking flow at the worst moment.

  Bottom Line

  Antigravity 2.0 is ambitious and directionally right. It bets that AI coding moves from assistant to autonomous agent,
  and it builds moats in multi-agent orchestration and cloud hosting. The trade-offs are real: a disruptive interface
  and a billing model that takes getting used to. For engineering teams, the smart move is to evaluate the SDK and
  plugin ecosystem now and plan where it fits in future automation pipelines.

   Antigravity https://antigravity.google/