Google Makes Interactions API the Main Gateway to Gemini Models and Agents

Google on June 22 made its Interactions API generally available and named it the primary way to build with Gemini models and agents. In a Google blog post, Google DeepMind's Ali Çevik and Philipp Schmid said the API now has a stable schema and is the default in Google's documentation, AI Studio and the Gemini API. For enterprise architects, the change matters because Google says its newest capabilities for long-running models and agents will increasingly ship only on this interface. Teams building on the older API now have a migration decision to make.

What is the Gemini Interactions API?

The Interactions API launched in public beta in December 2025. It is a single interface for two kinds of calls: inference against a model, using a model ID, and autonomous tasks handed to an agent, using an agent ID.

Google described its scope as server-side state, background execution, tool combination and multimodal generation. In practice, that includes:

• Background execution. Developers can set a flag on any call so the server runs the interaction asynchronously, which suits tasks that take minutes or longer.
• Stored interactions. Past interactions can be retrieved later, with 55-day retention on the paid tier.
• Managed agents. One API call provisions a remote Linux sandbox where an agent can reason, run code, browse the web and manage files. Google's Antigravity agent is the default, and teams can define custom agents with their own instructions, skills and data sources.
• Deep Research agents. Two new versions trade speed against depth, and they add collaborative planning, charts and grounding on images, PDFs and audio.
• Combined tools. Built-in tools such as Google Search and Google Maps can be mixed with a company's own functions in a single request.

The schema also replaces the familiar chat roles with typed steps, such as user input, model thought, function call and model output.

Does the old Gemini API still work?

Yes, for now. Google said the legacy generateContent API remains supported, and it published a migration guide plus a coding-agent skill to help developers move over. But the company was explicit that frontier features for long-running work will increasingly land "exclusively on the Interactions API." Integrations are already available from partners including LiteLLM, Eigent and Agno.

What does this mean for enterprise AI teams?

The launch is part of a broader shift by model providers from selling stateless text completions toward hosting stateful, long-running agent work on their own infrastructure. That brings convenience and new governance questions at the same time.

Data residency and retention. When a provider stores conversation state and agent steps on its servers, the 55-day retention window becomes a policy question, not just a technical detail. Legal and privacy teams should review what is stored, where, and how it can be deleted, especially for regulated data.

Sandboxed execution. A managed agent that runs code and browses the web inside a provider-hosted sandbox shifts part of the security perimeter to that provider. Security teams will want to know what network access the sandbox has and which enterprise systems an agent can reach through custom tools.

Lock-in. Typed steps, server-side state and provider-hosted agents make applications more capable, but also harder to move. Speakers at AI Connect events have warned about this repeatedly. "You don't want their SLAs to be your SLA," Joe McCarron, Head of Developer Advocacy at Skyflow, said of model providers during a AI Connect webinar on scaling AI architecture. He argued for observability and graceful degradation when a vendor's API goes down.

Earlier AI Connect panels made the same point from a strategy angle. SandboxAQ's Stefan Leichenauer, VP of Engineering, warned in a AI Connect expert panel webinar that a chosen vendor "may not even exist two years from now," and advised open frameworks and multivendor proofs of concept. The panel on future-proof AI infrastructure remains a useful checklist.

Cost controls. Google said Flex and Priority tiers let developers optimize for cost or latency, with Flex described as offering a 50% cost reduction. Background agents that run for long periods can consume a lot of tokens, so teams should decide upfront which workloads go to the cheaper tier and set budgets per agent.

How should architects respond?

A few practical steps make sense before adopting the new interface broadly:

• Inventory Gemini dependencies. Identify applications on the legacy API and decide which need the new agent features and which can stay put for now.
• Keep an abstraction layer. Route calls through an internal gateway or an open-source router, so a future switch between providers does not mean rewriting every application. AI Connect's coverage of enterprise agent platforms describes this pattern.
• Instrument from day one. Log agent steps, tool calls and token use in your own observability stack, not only in the provider's console.

The next milestone to watch is Gemini Omni, which Google listed as coming soon to the API. How quickly Google stops adding features to the older interface will determine how urgent the migration becomes for enterprise teams.

Frequently asked questions

What is Google's Interactions API?

It is Google's primary API for Gemini models and agents, generally available since June 22, 2026, supporting server-side state, background execution, combined tools and multimodal output.

Is the old Gemini generateContent API going away?

Google says it remains supported, but new capabilities for long-running models and agents will increasingly be available only on the Interactions API.

How long does Google retain interactions?

Google says past interactions are retrievable with 55-day retention on the paid tier.

Filed under: Agentic AI, Data & AI Infrastructure, AI Governance & Trust