Kevin Hou

Kevin Hou led product at Windsurf, where he helped bring agentic coding into the editor and worked on Cascade’s web-retrieval tools. His public explanations emphasize software engineering as an ongoing workflow: agents need to understand what a developer is already doing, use information beyond the repository, and continue partially completed tasks. In November 2025, he described joining Google DeepMind four months earlier and participating in his team’s Google Antigravity launch.[8][7][9]

In his November 13, 2024 Windsurf introduction, Hou credited the team with a September breakthrough in codebase understanding, tool use, and coordination with the user that became Cascade. He assembled the accompanying product demonstration. His explanation of the editor centered on inferring intent from typing, autocomplete, and codebase navigation, then using that context to suggest multi-cursor edits, identify the next file to change, or continue the developer’s existing task.[6]

At AI Engineer World’s Fair 2025, Hou described Windsurf’s organizing idea as “a shared timeline between the human and the AI.” In his keynote, that timeline included file reads and terminal commands, external information such as documentation and previous pull requests, and organizational practices that influence what counts as good code.[8]

Hou illustrated the broader workflow with a new dashboard page: customer requests in Slack inform planning in Google Docs and Jira, a designer works in Figma, and implementation is followed by pull requests, review, merging, and deployment. He argued that an agent needs access to the requirements and decisions surrounding the code, along with tools for acting beyond the editor. His proposed direction was background work that returns to the human for final approval.[8]

Hou’s January 17, 2025 technical account explained Cascade’s web retrieval as three distinct tools. Web Search finds candidate pages without reading their contents; URL Read retrieves a page and may divide it into chunks; View a Chunk reads selected sections. Short pages could enter the context directly, while long pages produced a summarized outline from which the agent could choose what to inspect.[7]

Site compatibility was a challenge Hou personally worked on. He considered outsourcing scraping but concluded that it would weaken the security posture and prevent access to documents within the user’s network. The implementation instead scraped locally, on the user’s machine and network. He also identified practical limits: client-rendered content could be missed, and paywalls, login pages, online editors, and JavaScript-heavy sites could cause problems.[7]

Hou wrote that agents should “mimic the behavior of humans when solving problems,” but his usage advice favored selective tool use. He recommended specific prompts and reliable documentation sources, and cautioned that forcing web search when no current information was needed could waste credits and slow the task.[7]

Hou connected the shared timeline to Windsurf’s SWE-1 training and evaluation priorities. He described two offline evaluations: an end-to-end task starting from an intent and codebase and ending with passing unit tests, and a conversational task testing assistance within an existing conversation or partially completed change. The second assessed helpfulness, efficiency, and correctness because real engineering work may begin with unfinished edits and require tools outside the editor.[8]

He also argued that product feedback should guide improvements across the whole system. Accept/reject actions and thumbs-up/down feedback could reveal missing tools, repeated workflows, and failure points. The response might involve model training, new tools, changes to the agentic harness, or better memories and checkpointing. His closing principle was that model, data, and application must develop together, with the application reflecting the behavior the model is meant to learn.[8]

In his November 19, 2025 Antigravity article, Hou recalled joining Google DeepMind four months earlier to continue building agentic coding tools. He described the team’s launch and produced a getting-started tutorial. His account emphasized artifacts, parallel background agents, and an agent-controlled browser.[9]

Hou explained artifacts as dynamic files that organize work, communicate with users, or provide context to another agent. As tasks become longer and changes more complex, he argued, a conversation stream becomes harder for a human to follow. Task lists, diagrams, screenshots, and recordings can instead present the information needed for review. Antigravity’s Agent Manager provided an interface for managing agents, reviewing those artifacts, and starting tasks.[9]

Writing in a personal capacity, Hou argued that agents need to verify their own work by clicking, scrolling, and using an application as a user would. The browser also supplies context for research, planning, and other engineering activity. He favored screenshots and recordings as a way to communicate completed work to the user.[9]

Before his agentic coding work, Hou worked on interfaces for other AI systems. His personal career account lists leadership of Nuro’s Autonomy Visualization team and the launch of Salesforce’s Einstein Designer, an AI generative UI product, at Dreamforce 2020. He holds a Computer Science B.S.E. from Princeton University. These earlier roles provide context for his combination of software engineering, design, and AI product work.[4][5]

Hou’s programming portfolio includes Ask Ava, a voice assistant using a BERT question-answering system to answer questions with podcast excerpts, and Slate Scheduler, which used a model of the user’s calendar to predict events and reduce typing. Both applied AI to concrete interface tasks.[5]