Mukund Sridhar is a software engineer who helped create Gemini Deep Research, Google’s AI assistant for investigating complex questions across the web and producing detailed, sourced reports. His work focuses on the systems required to make research agents persistent, transparent, and reliable across lengthy, multistep investigations.
Sridhar studied at the University of Massachusetts Amherst from 2013 to 2015 and developed interests in natural-language processing and machine learning. He contributed to technical reports introducing the original Gemini model family and Gemini 2.5. Working with product manager Aarush Selvan, he helped develop Deep Research from an early prototype that compared more than 20 summer camps around New York. After its release, he described the product’s development as a collaboration across modeling, inference infrastructure, and user experience.
How he builds research agents
- Long-running research workflows: Investigations involving repeated model calls and multiple services require durable state, recovery from intermediate failures, asynchronous execution, and notifications that let people resume on another device.
- Iterative planning under uncertainty: Agents must determine which questions can run in parallel, recognize incomplete answers, and redirect subsequent searches. Finding athletic standards for one competitive division, for example, leaves the original question unresolved when other divisions remain unexamined.
- Entity resolution across fragmented sources: Research systems must distinguish whether separate webpages describe the same organization or offering, then combine complementary details such as scuba-course requirements and prices without conflating distinct entities.
- Retrieval-augmented research notes: As browsing results and follow-up questions accumulate, agents can retain recent findings in their active context while storing older information in retrievable notes.
Sridhar’s account of Deep Research’s architecture frames useful research assistance as an engineering problem spanning planning, memory, failure recovery, and the unpredictable structure of the open web.
