Kevin Hou

Kevin Hou leads product engineering for Google Antigravity at Google DeepMind, building software-development environments where AI agents can write code, operate browsers, verify applications, and coordinate parallel tasks. Previously, he led product engineering at Codeium and Windsurf, helping move coding assistants beyond autocomplete toward sustained participation in the wider engineering workflow.

From autonomous vehicles to agent-first development

Hou studied computer science at Princeton University and worked on autonomous vehicles at Nuro before joining Codeium, the AI developer-tools business that emerged from Exafunction. His programming portfolio reflects a broader interest in full-stack software, product design, and machine learning; he also works in photography and woodworking. Earlier projects include async-post-message and Handwriting Font App, which converts handwritten characters into text suitable for physical reproduction.

At Codeium, Hou tackled a basic weakness of AI coding tools: useful changes usually depend on several related files, project-specific conventions, and external documentation, while conventional retrieval benchmarks often reward finding one relevant document. The team developed M-Query, which uses models to assess candidate codebase items and assemble richer context. Its evaluations paired pull requests and commits with the files they changed, measuring retrieval against realistic engineering work while accounting for latency and infrastructure costs. His account of repository-aware retrieval also explains why larger context windows and ordinary vector embeddings do not automatically solve the problem.

As Codeium evolved into Windsurf, Hou helped develop an agent-oriented editor organized around a shared human-agent timeline. File edits, searches, terminal commands, and agent actions contribute to the same working context, enabling an agent to continue unfinished tasks, infer dependencies, and remember project conventions. Potentially dangerous commands require approval. His Windsurf product architecture discussion frames persistent context as a prerequisite for collaboration, not merely a convenience.

Windsurf subsequently expanded into browser previews, GitHub pull requests, background agents, and integrations with tools including Notion, Linear, and Stripe. Hou introduced Windsurf Previews as a way to modify applications through direct interaction with their running interfaces. Windsurf’s SWE-1 model extended that approach to evaluation: assess complete engineering tasks and partially finished conversations, not code generation in isolation.

Hou helped launch Google Antigravity in November 2025, combining an AI editor, an agent-controlled browser, and Agent Manager for supervising concurrent work. The browser can retrieve authenticated context and test applications directly; implementation plans, diagrams, screenshots, recordings, and walkthroughs make an agent’s decisions and results reviewable.

• Repository-aware retrieval: Coding agents need multiple interconnected sources of context, evaluated against actual code changes rather than simplified search exercises.
• Shared human-agent context: Agents should track evolving files, terminal activity, documentation, and project preferences so they can continue work without repeatedly requesting the same information.
• Browser-based verification: A useful agent should inspect and interact with the running application, producing visible evidence of whether its changes work.
• Artifacts as an agent interface: Plans, task lists, diagrams, screenshots, and walkthroughs give developers practical checkpoints for reviewing, redirecting, and approving autonomous work. Hou’s explanation of Antigravity’s design connects these interfaces to parallel orchestration and collaboration between product engineers and model researchers.