Google's Interactions API Reaches GA: The New Standard for Gemini Agents
Google has officially launched the Interactions API in General Availability, making it the primary, stateful interface for Gemini models and agents.
The Interactions API is now the recommended interface for building with Gemini models and agents.
Google has officially announced the General Availability (GA) of its Interactions API, marking a major milestone in how developers build with Gemini models and agents. First introduced as a public beta in December 2025, the API has rapidly evolved from an experimental endpoint to Google’s primary interface for stateful, agentic workflows.
While the legacy generateContent API will remain fully supported for the foreseeable future, Google has made it clear: the Interactions API is now the default. For developers building next-generation applications, all frontier capabilities for long-running models and advanced agents will increasingly launch exclusively on this new architecture.
Why the Shift? An Agent-First Architecture
The transition to the Interactions API represents a fundamental shift in how applications communicate with AI. Instead of stateless, request-response loops that force developers to manually manage and resend conversation histories, the Interactions API centers around an Interaction resource that natively supports server-side state management.
- Server-Side State: By passing a
previous_interaction_id, developers can seamlessly continue conversations. This improves performance and drastically lowers costs by leveraging implicit context caching for conversation history. - Background Execution: Setting
background=Trueenables the server to run long, complex asynchronous tasks—ideal for deep reasoning or autonomous agents. - Unified Interface: Whether you are calling standard models like Gemini 2.5 Pro or specialized agents like Deep Research, the unified endpoint handles it all seamlessly.
"The Interactions API gets you there in a few lines of code. Pass a model ID for inference, an agent ID for autonomous tasks, set background=True for anything long-running."
Key Features Launching with GA
The jump to General Availability isn't just a stability update; it brings a massive suite of features targeting power users and complex agent orchestration.
Managed Agents and Deep Research
A single API call can now provision a remote Linux sandbox, granting agents the ability to execute code, browse the web, and manage files. Google is shipping the Antigravity agent as the default Managed Agent. Developers can also define custom agents with tailored instructions, skills, and data sources.
For more intensive reasoning tasks, the Deep Research agent has received massive upgrades. Developers can now choose between speed-optimized or depth-optimized versions. The agent also boasts collaborative planning capabilities, native chart and infographic generation, and robust multimodal grounding (combining images, PDFs, and audio).
From Roles to Steps
A major structural improvement is the simplified schema. The old, somewhat rigid "role" structure has been replaced by typed execution Steps. Every action in a session is now explicitly defined as a user_input, thought, function_call, or model_output. This provides unprecedented observability for debugging and UI rendering.
Cost Optimizations and Media Generation
To help scale production apps, the API introduces Flex and Priority tiers. The Flex tier is designed for latency-insensitive tasks and offers a massive 50% cost reduction. Additionally, the Interactions API fully supports Google's latest media generation tools, including Nano Banana 2 for image generation (grounded with Google Image Search), Lyria 3 for music generation, and multi-speaker TTS for highly expressive speech output.
Migration and Ecosystem Support
Google is pushing hard for ecosystem adoption. The Interactions API is supported out of the box in the latest versions of the official SDKs (Python google-genai version 1.55.0+ and JavaScript @google/genai version 1.33.0+). Third-party orchestrators and gateways like LiteLLM, Eigent, and Agno have already implemented support.
To assist developers relying on AI to write their code, Google released the gemini-interactions-api Skill. By connecting an assistant to the Gemini Docs MCP and injecting this skill, coding agents instantly learn the best-practice patterns for the new API, including streaming, structured output, and function calling.
Current Limitations
While the Interactions API is comprehensive, there are a few legacy features yet to be ported over. As of June 2026, the Batch API and Video metadata controls (for custom frame rates and clipping) are still restricted to generateContent. Additionally, Gemini 3 models do not yet support remote MCP, though Google states this is coming soon.
The Bottom Line
With its stable schema, stateful architecture, and robust support for the latest Gemini 3.1 models and agents, the Interactions API is the definitive path forward for AI development on Google Cloud. Developers looking to build agentic, long-running, and multimodal applications should begin planning their migrations today.
Enjoyed this?
Get more posts like this delivered to your inbox.
