Windsurf everywhere, doing everything, all at once - Kevin Hou, Windsurf

Key Concepts

Windsurf editor, AI-assisted software engineering, shared timeline, agentic workflows, context ingestion, meta-learning, software engineering model (Sui 1), data flywheel, frontier model, end-to-end task benchmark, conversational suite task benchmark.

Windsurf's Evolution and Philosophy

Kevin, Head of Product at Windsurf, discusses the evolution of Windsurf and its underlying philosophy. He emphasizes the rapid pace of innovation in the AI-assisted software engineering space. Windsurf has evolved significantly since its launch, moving beyond simple autocomplete to agentic workflows capable of complex tasks. The core philosophy revolves around a "shared timeline" between the human and the AI, allowing the AI to anticipate and assist the developer's workflow more effectively.

The Shared Timeline: From Autocomplete to Agentic Workflows

The presentation outlines the evolution of AI's role in software development:

• Pre-Autocomplete: AI had minimal impact, requiring manual coding.
• Early Autocomplete (e.g., Copilot): AI assists with small code snippets, accelerating minor edits.
• Windsurf Era (Late 2024): AI can perform more complex tasks, including editing multiple files, conducting background research, and executing terminal commands.

Windsurf aims to transform software creation by extending this timeline beyond the IDE.

Context Ingestion: Windsurf Being "Everywhere"

Windsurf aims to be ubiquitous by ingesting context from all sources a developer uses. These sources are categorized as:

• Coding-Related: File reads, terminal command history, open editor tabs.
• External Sources: GitHub commit history, pull requests, online searches, documentation.
• Meta-Learning: Organizational best practices, engineering preferences that define "good code."

Example: Building a new page on a data visualization dashboard involves Slack for context, Google Docs for planning, Jira for ticket tracking, Figma for design, and finally, code writing in Windsurf.

Windsurf connects to various MCP (Multi-Context Platform) services like Notion, Linear, and Stripe to access information.

Beyond Reading: AI Taking Action

Windsurf's goal is not just to read context but also to do everything a software engineer can. This includes:

• Interacting with third-party services.
• Provisioning API keys.
• Writing design documents and PRDs.
• Wireframing and testing.

Example: Building a web app involves running codebase-relevant terminal commands, using Windsurf browser previews for visual iteration, opening pull requests via GitHub MCP (using context from other PRs), and deploying to Netlify with one click. Windsurf Reviews can automatically leave comments and suggest changes asynchronously.

The Future: Always-On, Parallel AI Assistance

The vision is for Windsurf to be constantly working in the background, even when the user isn't actively coding. The goal is to shift from 80-90% agent and 10-20% human interaction to 99% agent and 1% human (final approval). This enables coding "anytime," including via voice activation.

Sui 1: Windsurf's Software Engineering Model

Windsurf developed its own software engineering model, Sui 1, because off-the-shelf frontier models weren't optimized for the messy, real-world workflows of software development. Sui 1 is trained specifically for software engineering workflows, not just code generation.

Evaluation Benchmarks for Sui 1

Sui 1 is evaluated using two main offline benchmarks:

1. End-to-End Task Benchmark: Measures the model's ability to complete pull requests from start to finish, passing all unit tests.
2. Conversational Suite Task Benchmark: Assesses how well the model assists in existing user conversations or partially completed tasks, reflecting Windsurf's "mid-timeline" assistance approach.

The blended score considers helpfulness, efficiency, and correctness. Sui 1 achieves near-frontier model results with fewer resources.

The Data Flywheel

Windsurf leverages a data flywheel:

1. Ship: Release the best product.
2. Use: Developers (and non-developers) use the product, improving their skills.
3. Find the Frontier: Users provide feedback (thumbs up/down, accept/reject), identifying missing tools and workflow bottlenecks.
4. Build at the Frontier: Windsurf trains better models, builds more tools, and improves its agentic harness.

This cycle repeats continuously.

Conclusion

Windsurf is striving to revolutionize software development by creating an AI-powered assistant that understands and participates in the entire software engineering timeline. By ingesting context from various sources, taking action on behalf of the developer, and continuously learning from user feedback, Windsurf aims to make software development more efficient and accessible. The development of Sui 1 demonstrates the company's commitment to building specialized AI models tailored to the unique challenges of software engineering. The ultimate goal is to create an "always-on" AI assistant that empowers developers to code anytime, anywhere.