Google Antigravity 2.0: Managed Agents in Gemini API — Interactions, CLI Migration & Routing (2026)

At Google I/O 2026, Google did not just ship a new model. It restructured the entire developer surface around agent orchestration — and in doing so, quietly raised the stakes for every team running multi-provider AI infrastructure.

The most consequential change for engineering teams is not Gemini 3.5 Flash (already covered here). It is Managed Agents in the Gemini API: a single-call primitive that spins up a fully isolated Linux environment, lets an agent reason and use tools, and persists that environment state across multi-turn sessions. The companion platform — Google Antigravity 2.0 — replaces the existing Gemini CLI, introduces a standalone agent-first desktop and Antigravity SDK, and signals that Google is now competing in the same architectural space as multi-agent orchestration layers like OpenAI Codex and Claude Code.

What changed

Google I/O 2026 (May 19–20) shipped five interrelated pieces:

1. Managed Agents via the Interactions API. A single API call creates an agent running inside an isolated Linux environment. The agent reasons, calls tools, and executes code with full persistent state across follow-up calls. The model underneath is Gemini 3.5 Flash by default; the orchestration harness is the same one Google uses internally. Available now in the Gemini API and Google AI Studio.

2. Antigravity 2.0 desktop app. A standalone application — not an IDE plugin — built around parallel agent execution, dynamic subagents, scheduled background tasks, and voice commands. The scheduling capability is notable: tasks can trigger agents automatically without a human prompt, effectively converting the tool from a single-turn assistant into a persistent automation pipeline.

3. Antigravity CLI replaces Gemini CLI. Gemini CLI is deprecated. The Antigravity CLI preserves Gemini CLI's Agent Skills, Hooks, Subagents, and Extensions (now called Antigravity plugins) but runs on the shared Antigravity agent harness. Migration is documented but not optional — the old CLI has a sunset date.

4. Antigravity SDK. Programmatic access to the same agent harness that powers Google's products. Developers can define custom agent behaviors and self-host on their own infrastructure. This is the path for teams wanting Antigravity-style agents without being locked into Google's managed execution plane.

5. Gemini Enterprise Agent Platform. The enterprise deployment path, connecting Antigravity to Google Cloud projects and targeting teams that need agents running inside their existing cloud governance perimeter.

Why it matters for AI engineering teams

The Managed Agents feature is the sharpest change for teams that currently call the Gemini API via OpenAI-compatible endpoints or standard chat completions. The Interactions API for Managed Agents is not an OpenAI-compatible surface. It is a distinct protocol: you invoke an agent ID (antigravity-preview-05-2026), pass unstructured input, and receive an agentic response from a stateful environment. That means:

• Existing routing proxies that forward /v1/chat/completions to Gemini will not route Managed Agent calls. The API shape is different.
• Cost model is different. Managed Agents bill against agent interaction sessions, not just token counts. Your current per-token cost accounting may not capture this correctly.
• Session state creates per-session stickiness. If an agent persists an environment across turns, routing load-balancing across provider instances must account for session affinity. A naive round-robin will break resumed sessions.
• Gemini CLI → Antigravity CLI migration affects developer toolchains. Teams using Gemini CLI in their local or CI workflows need to plan the migration. Extensions (plugins) that wrap Gemini CLI behavior will need to be ported to Antigravity plugin format.

The Antigravity SDK is the more router-friendly surface: it is self-hosted, Gemini-optimized but potentially adaptable, and lets teams keep orchestration under their own control. But if you want the Managed Agents execution environment, you are calling Google's cloud — not your own routing layer.

The router/operator angle

This announcement introduces a real architectural fork in how teams consume Gemini:

Path A: Standard Gemini API via OpenAI-compatible proxy. You call /v1/chat/completions (or Gemini native), your routing layer handles fallback, cost capping, load balancing. This path is unaffected by Antigravity unless you migrate.

Path B: Managed Agents via Interactions API. Google manages execution, persistence, and tool access. Your routing layer is effectively bypassed for agent calls. You trade orchestration control for Google-managed reliability and environment persistence.

Path C: Antigravity SDK, self-hosted. Custom agent harness on your infra, Gemini models underneath. More integration work, but your routing and governance layer can wrap the SDK calls.

The key operator question is: for which workloads do you surrender the orchestration layer to Google? The Managed Agents proposition is compelling for ephemeral, tool-heavy, stateful tasks (research sweeps, multi-step code generation, file-system tasks). But accepting it means accepting that those calls are not routable through your standard gateway, do not produce standard token-usage telemetry, and cannot fail over to a non-Google provider without rewriting the client.

Separately, the Gemini CLI deprecation creates an urgent migration window for any team with Gemini CLI in their toolchain. The Antigravity CLI is compatible at the feature level, but the rename, plugin format change, and new harness mean that automated scripts invoking gemini commands will need updating. Budget time for this before the sunset date.

What TheRouter users should watch or try

• Audit whether your Gemini integration uses chat completions vs. agentic endpoints. If you are calling standard /v1/chat/completions through a routing proxy, the Antigravity launch does not break that path today. However, if Managed Agents provide the features your team wants (persistent environment, tool use, multi-turn state), you will need to decide whether to call them directly or wrap them.

• Watch the Interactions API pricing carefully. Session-based agent billing is a different accounting model than per-token billing. Before committing to Managed Agents in production workflows, verify how usage will appear in your billing reconciliation and whether your cost-control tooling can interpret it.

• Plan the Gemini CLI migration. If CI pipelines, developer scripts, or agent scaffolds reference the gemini CLI, schedule migration to Antigravity CLI before the shutdown date. Test Extensions → Antigravity plugin compatibility, especially custom tool definitions.

• Evaluate the Antigravity SDK for self-hosted scenarios. If you want Antigravity-style parallel agents but need to keep orchestration inside your own routing perimeter, the SDK is the right surface. It is Gemini-optimized but architecturally separable from Google's managed execution plane.

TheRouter routes standard OpenAI-compatible requests across providers. Managed Agent sessions via the Interactions API are outside that scope by design — they carry session state that requires provider-side persistence. The router/operator decision is whether to treat Managed Agents as a separate, provider-managed service (and simply track its costs separately), or to invest in the Antigravity SDK path to preserve in-house orchestration control.