Google has elevated its Interactions API to General Availability, cementing it as the main channel for working with Gemini models and agents. In a move that mirrors the industry's shift toward more autonomous and portable AI, the API consolidates inference and agent orchestration into a single endpoint, doing away with legacy role structures in favor of a streamlined, step-based schema.
A Unified Gateway for Models and Agents
Instead of juggling separate interfaces for model inference and agent tasks, developers now pass either a model ID or an agent ID to the same endpoint. This simplicity extends to managed agents, where a single API call provisions a remote Linux sandbox. The built-in Antigravity agent can reason, run code, browse the web, and handle files, while custom agents can be defined with specific instructions, skills, and data sources.
The API also introduces background execution – just set background=True to run interactions asynchronously. This is a critical feature for portable use cases, where long-running tasks can complete server‑side without requiring a persistent client connection.
On the tooling front, developers can mix built-in tools like Google Search and Google Maps with custom functions, and tool results can now carry images alongside text. Multimodal generation gets a boost too: image generation via Nano Banana 2 with Google Image Search grounding, music with Lyria 3, and multi‑speaker text‑to‑speech.
What's New with GA
The GA release delivers a stable schema and several enhancements that developers have been asking for:
- Managed agents with sandboxed execution, default Antigravity agent, and customization options.
- Background execution for any interaction, freeing clients from waiting.
- Step‑based schema replacing the old role system; every action – user input, thought, function call, model output – is a typed step.
- Cost and latency controls: Choose between Flex (50% cost reduction) and Priority tiers.
- Error localization that pinpoints the exact field causing a failure.
- 55‑day interaction retention on the paid tier for auditing and replay.
- Partner ecosystem integrations with LiteLLM, Eigent, and Agno already supporting the Interactions API.
Google is not abandoning its legacy generateContent API, which will continue to receive core Gemini models. However, new frontier capabilities – especially those involving stateful, long‑running agents – will land exclusively on the Interactions API. A migration guide is available for teams transitioning at their own pace.
The Interactions API was designed from the ground up for stateful, agentic workflows – precisely the kind of logic that portable AI assistants demand. Background execution and managed sandboxes mean an agent can be launched from a mobile device and left to work autonomously, returning results when needed. The unified endpoint reduces integration complexity, while the Flex tier cuts costs for applications that must run efficiently on edge hardware.
Multimodal tool outputs (images and text together) and expressive speech synthesis make interactions more natural in hands‑free or glanceable scenarios. And because the API is now the default across Google AI Studio, Gemini API, and third‑party SDKs, developers building portable AI experiences can adopt a single, future‑proof interface. As Google puts it, the Interactions API is "our primary API for interacting with Gemini models and agents" – and for anyone targeting the next wave of portable, always‑available AI, that's a signal worth heeding.