Google makes Interactions API the default way to build with Gemini agents

#ai #google #gemini #agents

Google makes Interactions API the default way to build with Gemini agents

Google has moved its Interactions API to general availability and says it is now the primary API for working with Gemini models and agents.

This matters now because Google is not just adding another endpoint. It is telling developers that the agent-shaped API is the default path for new Gemini work, while the older generateContent API remains supported but may not get every frontier agent capability first.

What Google announced

Google says the Interactions API is now generally available after a public beta that began in December 2025. The GA release brings a stable schema and makes Interactions API the default across Google AI Studio, the Gemini API documentation, and new code snippets.

The important builder-facing pieces are:

One endpoint for models and agents. Developers can pass a model ID for ordinary inference or an agent ID for longer-running autonomous tasks.
Managed Agents. A single API call can provision a remote Linux sandbox where an agent can reason, run code, browse the web, and manage files. Google says the Antigravity agent is the default, and custom agents can be defined with instructions, skills, and data sources.
Background execution. Setting background=True lets the server run long tasks asynchronously.
Tool mixing. Built-in tools such as Google Search and Google Maps can be combined with custom functions in one request, and tool results can return images as well as text.
Deep Research upgrades. Google lists speed/depth agent variants, collaborative planning, native charts and infographics, and multimodal grounding with images, PDFs, and audio.
Media generation hooks. The post names image generation with Nano Banana 2 and Google Image Search grounding, music with Lyria 3, and multi-speaker text-to-speech.
Cost controls. Flex and Priority tiers let teams choose between lower cost and lower latency; Google says Flex offers a 50% cost reduction.
State retention. Paid-tier users can retrieve past interactions for 55 days.

Google is also moving the schema away from the old role-based message format. In the new model, each action is a typed step — user input, thought, function call, model output, and so on.

Why builders should care

If you are building against Gemini, this changes the default architecture decision.

For simple chat or extraction calls, generateContent is still supported. But if your product needs long-running tasks, agent state, tool calls, web or map grounding, multimodal outputs, or remote execution, Google is clearly steering you toward Interactions API.

The bigger signal is about where new capabilities will land first. Google says the legacy API will continue receiving new mainline Gemini models "for the foreseeable future," but it expects frontier capabilities for long-running models and agents to increasingly arrive exclusively on Interactions API.

That is the line teams should pay attention to. If you are starting a new Gemini integration today, the risk is not that the old API breaks tomorrow. The risk is building on the path that gets the new agent features later, or not at all.

Practical next steps

For engineering teams, I would treat this as a migration-planning item rather than a panic rewrite.

New Gemini agent projects should start on Interactions API.
Existing generateContent apps should keep running, but teams should audit which workflows would benefit from background execution, managed agents, or built-in tool combinations.
Wrapper libraries and internal SDKs should be checked for support. Google names LiteLLM, Eigent, and Agno as early supported partners.
Cost-sensitive workloads should test the Flex tier rather than assuming Gemini agent calls have to run at the same latency/cost profile as interactive chat.

Caveats

Google says Gemini Omni support is coming soon, not fully here today. Also, "Managed Agents" means more of the runtime is sitting on Google's side, so teams handling sensitive code or regulated data should review sandbox behavior, retention, logging, and data controls before moving production workflows.

The old API is not deprecated in this announcement. Google is saying where the platform is going, not turning off the current path.