Google Releases Gemini Deep Research Agent API

Google has released the Gemini Deep Research Agent as an API. The Deep Research feature, previously only available through the Google AI Studio web UI, can now be integrated directly into developers' own apps.

The core interface is the new Interactions API. Rather than returning immediately like a typical generate_content call, it operates as an asynchronous task that runs in the background for several minutes. After sending a request, developers can check progress via periodic polling or via streaming.

Two models are provided.

• deep-research-preview-04-2026: focused on speed and efficiency
• deep-research-max-preview-04-2026: focused on comprehensiveness and automated synthesis

Notable features include the following.

• Collaborative Planning: instead of executing immediately, the agent first presents a plan, and once the user revises and approves it, execution proceeds.
• Automatic chart/infographic generation: setting visualization="auto" automatically generates visualizations and returns them as base64 images.
• MCP server integration: external tools such as financial data sources can be attached to expand the scope of research.
• Default toolset: Google Search, URL content reading, and code execution are provided by default.
• Multimodal input: in addition to text, images, PDFs, and audio can be included as research context.
• Real-time streaming and thinking summaries: progress can be received via streaming, and setting thinking_summaries="auto" allows viewing intermediate reasoning summaries.

The example code shows a flow where a task is started with background=True and then completion is checked via client.interactions.get(...). For collaborative planning, collaborative_planning=True is used to receive only the plan, and it must be switched to False at the end to start the actual research.

The meaningful change here is that the deep research workflow has come down to the API level without requiring an agent framework. In other words, long-running research tasks—including planning, search, synthesis, and visualization—can now be built directly into applications. However, since this presupposes delays on the order of several minutes, how to design the asynchronous task and status-waiting flow within the product UX becomes a key consideration.