The Inventors of Deep Research — Inside Gemini Deep Research at Google DeepMind
Aarush Selvan (PM) and Mukund Sridhar (Tech Lead) from Google DeepMind's Gemini Deep Research team join Latent Space to walk through how they built the product that, by their account, originated the entire "Deep Research" agent category — before OpenAI's Deep Research launched on Feb 2 and a wave of clones followed. They explain why Deep Research is an agentic, ~5-minute web research workflow (not just LLM search), and unpack the design choices: an editable, transparent research plan shown upfront; real-time visibility into the websites being read; and a side-by-side artifact + chat layout so users can keep iterating without losing context. On the technical side, the team describes post-training a special version of Gemini 1.5 Pro (rather than using the base model) so the agent can do iterative, parallelizable planning across two core tools (search and deeper webpage reading), then enter an analysis mode that reasons over inconsistencies, drafts an outline, and self-critiques before finalizing the report. They discuss long-context vs RAG trade-offs (keeping recent research in-context for follow-ups, falling back to RAG for older turns), HTML-to-markdown handling, vision deferred due to latency cost, and an in-house asynchronous orchestration platform built to handle multi-minute, retryable jobs with durable state. The conversation also covers evaluation (high-entropy outputs, a mix of auto-metrics like research-plan length and iteration counts plus heavy human evals against an ontology of use cases spanning broad/shallow to narrow/deep), the counterintuitive finding that users reward longer waits as evidence of effort, and where the product is headed: personalization, multimodal in/out and generative UI, and complementing the open web with private documents and subscriptions. They note that next-gen reasoning/thinking models are being explored but introduce new tension between leveraging internal model memory and grounding in sources.
Original sources
Podcast Latent Space
Episode:The Inventors of Deep Research ↗
YouTube Latent Space
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.