/v1beta/interactions spins up an isolated Linux sandbox, runs your agent, and returns a stateful environment ID that persists across sessions. No database, no queue, no state machine on your end. The launch is paired with ADK 2.0 GA across Python, TypeScript, Go, Java, and Kotlin.
The Infrastructure Problem, Solved
Deploying a production AI agent typically means wrangling VMs, Docker containers, a state database, queuing infrastructure, and an orchestration layer. Google's pitch with the Managed Agents API is straightforward: make the infrastructure problem disappear into an HTTP call.
The core mechanism is the new Interactions API (POST /v1beta/interactions). Pass an agent ID and a prompt, and Google provisions an isolated Linux container, runs the agent inside it, and returns the result. The agent has full Bash terminal access, can install packages, and can write files.
Stateful by Design
What separates the Managed Agents API from a simple code execution sandbox is persistence. Each interaction returns an environment_id. Pass that ID on your next call and the agent resumes in the exact same environment — same filesystem, same installed packages, same working state — without any database or queue on your end.
The sandbox has a 7-day TTL. Network isolation is off by default; external access requires an explicit allowlist. Paid-tier interactions are retained for 55 days; free tier, 1 day.
This design enables a class of long-running agents that couldn't previously exist without significant infrastructure investment:
- Multi-day HR onboarding coordinators waiting for human signatures
- Agents that autonomously build and configure operating systems
- Creative agents generating complex 3D scenes in Blender over several days
You need the
google-genai Python package at version 1.55.0 or later. The default managed agent — antigravity-preview-05-2026 — runs on Gemini 3.5 Flash and uses the same runtime powering Google's own first-party products. Network isolation is off by default; enable external access via an explicit allowlist in your interaction request.ADK 2.0: From Hierarchical to Graph-Based
Google also shipped ADK 2.0 to general availability at Google I/O 2026. The headline change: replacing the old hierarchical executor with a graph-based workflow engine. This gives developers a continuous slider between dynamic model-led reasoning and strict deterministic pipelines.
| Feature | ADK 2.0 |
|---|---|
| Workflow engine | Graph-based (replaces hierarchical executor) |
| Execution style | Dynamic reasoning ↔ deterministic pipelines |
| Multi-agent support | Collaborative Workflows with parallel/sequential sub-agents |
| Languages (GA) | Python, TypeScript, Go, Java, Kotlin |
| CLI | agy command (replaces Gemini CLI) |
| IDE | Antigravity 2.0 (VS Code-based, up to 5 parallel agents) |
Antigravity 2.0 and the Full Stack
The launch includes Antigravity 2.0 — a full desktop IDE (VS Code-based) with a Manager View that orchestrates up to five parallel agents simultaneously. The agy CLI replaces the old Gemini CLI. Pricing runs from free preview up to $700/month for the highest quota tier.
Google positions this as its complete Gemini Enterprise Agent Platform: ADK for building, Agent CLI for quick deployment, Antigravity for IDE-based development, and Managed Agents for hosting — all connected through the Interactions API.
— Gemini Managed Agents API: Public preview (built-in Antigravity & Deep Research agents)
— ADK 2.0: GA in Python, TypeScript, Go, Java, Kotlin
— Antigravity 2.0 IDE: Public preview
— Pricing: Free preview up to $700/month (highest quota tier)
— Custom agents: Build with ADK and host on your own infrastructure using the SDK
How It Compares to OpenAI and Anthropic
Google, OpenAI (Codex), and Anthropic (Claude Code) are all converging on cloud-hosted sandbox architectures. The key differentiator for Gemini Managed Agents is the isolation-by-default posture: agents run inside Google's infrastructure, isolated from external networks unless you explicitly open them. This optimizes for a different threat model than Claude Code, which runs locally with full filesystem and network access. Neither approach is universally better — they target different deployment contexts and trust requirements.
Key Takeaways
- Gemini Managed Agents API: Single POST call provisions a stateful Linux sandbox agent — no VMs, Docker, or state management required
- environment_id enables session resumption with same filesystem state across calls
- ADK 2.0 GA: graph-based workflow engine, Python/TS/Go/Java/Kotlin supported
- Antigravity 2.0 IDE: VS Code-based with Manager View for up to 5 parallel agents
- Long-running agents (days to weeks) officially supported with durable state checkpoints and event-driven dormancy