Most of the coverage I saw about Antigravity 2.0 focused on what it isn’t. Yes, there’s a new desktop IDE. Yes, the Gemini 3.5 Flash benchmarks are interesting. But if you’re making decisions about your team’s tooling, neither of those two things is the headline. The headline is a single API call that delivers you a Linux box running with an agent inside — and a migration deadline that’s going to break a lot of pipelines on June 18th.
Let me separate the signal from the noise of launch day.
What actually launched
Antigravity 2.0 is no longer a single product. Google consolidated its agent efforts into a single harness exposed through five surfaces: the desktop IDE, a new terminal CLI invoked with agy, a Python SDK, the Managed Agents API, and an enterprise path through Google Cloud. Everything runs on Gemini 3.5 Flash and — this is the part that matters — everything shares the same agent harness underneath. One behavior model, five places to invoke it.
That consolidation is the strategic move. The individual surfaces are interesting, but what deserves your attention is the Managed Agents API, because it changes what “building an agent” costs you in terms of infrastructure.
The Managed Agents API: the only genuinely new primitive
Here’s the capability in one sentence: a call to the Interactions API provisions an isolated, ephemeral Linux sandbox where an agent can reason, write and execute code, manage files, and navigate the web — and the environment persists between calls, so you can pick up a session with all state intact.
If you’ve ever built agent infrastructure, you know what that sentence is really saying. The work that used to mean spinning up sandboxes, wiring up tool-calling, managing isolation, and handling the lifecycle of a code execution environment — Google now does all that behind a single function call. You pass an agent identifier, specify a remote environment, describe the task, and get back a result produced inside a real sandbox you didn’t have to provision.
State persistence is the detail I’d underline. Previous code-interpreter-like functions restarted the container on each call, so anything you’d installed with pip install, any file you’d written, any half-open interpreter — gone. Here, if you pass the same environment ID in a follow-up call, the agent sees the previous /workspace/ exactly as it left it. That’s the difference between a stateless function and an actual working environment.
The convergence nobody at Google is going to say out loud
These agents are configured with files. Behavior instructions go in an AGENTS.md at the root. Custom skills live as Markdown files under .agents/skills/, each in its own subfolder with a SKILL.md. Reference data lives in a workspace directory.
If you’ve used Claude Code, that structure is going to feel extremely familiar — because it’s the same pattern. The filesystem-native, Markdown-defined agent configuration that Anthropic popularized is quietly becoming a cross-platform convention, and Google adopting it wholesale is the clearest signal yet that that’s where agent tooling standardization is happening. For your team, the practical advantage is real: the mental model you built for one tool now transfers. That’s rare in this space, and worth capitalizing on.
The part that’s going to hurt: the Gemini CLI dies on June 18th
Now for what’s at stake. On June 18th, 2026, the Gemini CLI and Gemini Code Assist extensions for IDEs stop serving requests for Google AI Pro, Ultra, and free Gemini Code Assist accounts for individuals. It’s not degraded service. It’s not a warning banner. The gemini command stops returning responses. Any shell script, CI pipeline, GitHub Action, or cron job that calls it breaks on that date.
Be precise about who this hits, because scope matters for your migration plan:
- Affected: Pro, Ultra, and free Gemini Code Assist for individuals. Gemini Code Assist for GitHub also stops accepting new installations on June 18th, with requests ceasing in the weeks following.
- Not affected: Paid Gemini API keys keep working indefinitely. Those with Gemini Code Assist Standard and Enterprise licenses keep access.
The replacement is the agy CLI — a rewrite in Go, faster cold starts, background async workflows, and maintains Agent Skills, Hooks, Subagents, and Extensions (now called Antigravity plugins). The migration itself is short for most setups. The reason to do it now rather than on the 17th is that the longer you stay with the old tool, the more fragile your scripts become, and Google openly stated that agy doesn’t have 1:1 feature parity at launch — so you want time to test, not surprises on deadline day.
Two things to see clearly
First, the successor is closed source. The Gemini CLI was released under Apache 2.0 with an open contribution model; agy is a proprietary binary, and Google hasn’t committed to opening it. Whether that matters depends on whether your decision to adopt it relied on auditability or the ability to fork. For many teams it won’t change the math. For some it will, and those teams should know it now.
Second, and more important operationally: the Managed Agents API and the Antigravity agent are in preview. Google’s own language is that this is Pre-GA software — no SLA, not intended for production use or sensitive data, and the API shape can change. That doesn’t make it useless; it makes it excellent for prototypes, internal tooling, and demos, and a bad choice for anything that bears weight until it reaches GA. Build with this, learn from this, but don’t put the weight of your production workload on a preview primitive.
What I’d actually do
If you’re running any terminal automation against Gemini CLI under an affected tier, treat the migration as this week’s work, not this month’s. Install agy, port your skills and your MCP config, and test before June 18th — the rollback window is free until then.
And if you’re building agents, the Managed Agents API deserves a real afternoon of experimentation. The infrastructure abstraction is the most compelling I’ve seen from a major provider, and the AGENTS.md convention means what you learn isn’t locked to a single vendor. Just keep it in the prototypes column until Google removes the preview label.