New agent APIs and research agents (Gemini / Interactions / TEN)

DeepMind released Gemini Deep Research Agent built on Gemini 3 Pro and introduced a unified Interactions API for models and agents. - Gemini Deep Research Agent autonomously plans deep research tasks, combines private docs with web data, and returns citation-backed reports. - Agent scored SOTA on proprietary benchmarks: DeepSearchQA 66.1%, Humanity's Last Exam 46.4%, BrowseComp 59.2%. - Interactions API provides a single endpoint for models and agents, offering developer controls for report structure, JSON outputs, and tool orchestration. - Google released the agent to developers via Google AI Studio and open-sourced a new test set for deep research evaluation. - TEN Framework introduced an open-source industrial agent runtime for real-time multimodal agents with TEN Designer visual flow, polyglot runtime, and edge-cloud orchestration for audio/video + LLM pipelines. - TEN addresses latency, glue-code, multi-language integration, and real-time interruptibility for voice/video assistants.
Key insights
Gemini Deep Research Agent aims to be an autonomous research assistant
DeepMind launched Gemini Deep Research Agent based on Gemini 3 Pro, designed for autonomous deep research: it plans searches, reads results, identifies gaps, and iterates targeted queries while integrating private documents and web sources. The agent produces reportable outputs with verifiable citations and is trained with multi-step RL to reduce hallucinations and improve factuality for complex investigative tasks.
Interactions API unifies model and agent integration for developers
DeepMind introduced the Interactions API—a single interface to work with both models and agent workflows—allowing developers to control report structure, request machine-readable JSON, and orchestrate multi-step agent tasks. The API is the integration point for Gemini Deep Research and is available via Google AI Studio with developer keys.
Agent benchmarks show state-of-the-art deep research performance
DeepMind released a new DeepSearchQA benchmark and reported top scores for Gemini Deep Research Agent: DeepSearchQA 66.1%, Humanity's Last Exam 46.4%, and BrowseComp 59.2%, claiming leading performance on tasks requiring causal-chaining, deep integration, and hard-to-locate information retrieval.
TEN Framework targets real-time multimodal agent engineering
TEN Framework provides an open-source runtime and visual designer (TEN Designer) for real-time multimodal agents handling audio, video, ASR, LLM, and TTS. It supplies a polyglot runtime supporting C++ for low-level media, Python for LLM logic, and Go for high-concurrency networking, plus edge-cloud partitioning to run lightweight models at the edge and heavy inference in cloud, addressing glue-code complexity and latency.