Taking Agents from 90 to 100 - Kevin Hou, Windsurf head of product engineering

Key Concepts

• Agentic Editor: An editor powered by AI agents that assist developers in coding tasks.
• Trajectories: A unified timeline of user actions and agent actions within the editor, allowing the agent to understand the user's context and intentions.
• Meta Learning: The ability of the agent to learn and adapt to a user's coding style, preferences, and organizational guidelines over time.
• Scale with Intelligence: Designing the product to leverage improvements in AI models, rather than relying on fixed infrastructure and rules.
• Cascade: The AI agent within Windsurf.

Windsurf: The First AI Agent Powered Editor

Introduction

Kevin, from the product engineering team at Codium, introduces Windsurf, an AI agent-powered editor. He emphasizes the belief that agents are the future of software development and highlights Windsurf's role in pushing the boundaries of this technology.

A Trip Down Memory Lane

• 2022: The Rise of Autocomplete: Kevin revisits the emergence of GitHub Copilot and Codium's initial autocomplete product, which gained millions of users.
• The Vision for the Future: Codium anticipated the evolution of AI models and aimed to build the best possible experience for developers, leading to the exploration of agents.
• 2025: The Age of Agents: Kevin asserts that 2025 marks the year where the power of agents in software development is widely recognized, with Windsurf leading the charge.

Vibe Coding with Windsurf

• Demo: A demonstration of Windsurf building a Python web scraper, showcasing features like dependency installation, code suggestions, and user-friendly accept/reject options.
• Key Features: Documentation lookup, web search, codebase awareness, commit message generation, and image support.
• Mission: To keep developers in the flow and unlock their limitless potential by handling grunt work and enabling them to focus on high-value tasks.

Input and Output

• Goal: To minimize explicit user input while maximizing the quality and readiness of the generated code.
• Strategies: Reducing human involvement through background research, predicting next steps, and making decisions on the user's behalf.
• Results: Windsurf has generated 4.5 billion lines of code in three months.

Principle 1: Trajectories - Reading Your Mind

• Deep Integration: Windsurf's agent is deeply integrated into the editor, understanding user actions and executing tasks on their behalf.
• "Continue My Work" Feature: The agent can continue the user's current task, potentially generating a full PR or commit.
• Terminal Execution Mode: The agent uses the LLM to determine the safety of terminal commands, prompting the user for confirmation when necessary.
• Unified Timeline: A shared timeline of user and agent actions, ensuring that the agent has an up-to-date understanding of the file state.
• Example 1: Adding a New Function: The agent suggests changes to other files and runs terminal commands based on the context of the new function.
• Example 2: Terminal Integration: The agent recognizes newly installed packages and suggests implementing them in the project.
• No Copy-Paste Future: The vision of a future where users don't need to copy-paste information between different tools.
• Ubiquitous Agent: The agent should be able to anticipate the user's needs and perform tasks automatically, such as writing unit tests or performing codebase-wide refactors.

Principle 2: Meta Learning - Adapting to You

• Inferred Understanding: The agent learns and remembers the user's coding style, preferences, and organizational guidelines.
• Autogenerated Memories: The agent builds a memory bank of user actions and preferences, such as preferred versions of libraries.
• Custom MCP Servers: Users can plug in their favorite tools to adapt to their workflow.
• Command Whitelisting/Blacklisting: Users can control which terminal commands the agent is allowed to run automatically.
• Explicit vs. Inferred Context: The goal is for the agent to infer context from the codebase and usage patterns, rather than requiring explicit instructions.
• Example: Architecture Overview: The agent remembers the project's purpose and available endpoints based on a brief overview.
• Documentation Auto-Learning: The agent automatically looks up documentation for the packages being used in the project.
• Rules File as a Crutch: The belief that most rules files will be unnecessary by the end of 2025, as the agent will be able to infer the necessary information.

Principle 3: Scale with Intelligence

• Leveraging Model Improvements: The product is designed to improve as AI models get better, rather than relying on fixed infrastructure.
• Deleting Chat: The chat interface was replaced with an agent-only interface (Cascade) to improve the quality of interactions.
• Dynamic Context Inference: The agent can dynamically infer the relationships between bits of code and documents, reducing the need for explicit "@" mentions.
• Web Search: The agent reads the web like a human, using the LLM to decide which search results to read and what parts of the page to focus on.
• Unsupervised Work: The agent will be able to perform more tasks autonomously, such as generating full PRs and reading complex documentation.

Conclusion

• The Engine is Key: The underlying engine of Windsurf is the secret sauce, enabling a new world of automated software development.
• High Adoption Rate: 90% of the code written by Windsurf users is generated by Cascade.
• Empowering Developers: The goal is to equip every software engineer with the best tools, which are agents.

Call to Action

• Download Windsurf and start using the magic today.
• Connect with Kevin on Twitter.