AI Development and Agent Engineering Practices and Tools

OpenAI guidelines, Cursor 2.1, and practical guides drive the transformation towards agent-led development.
Key insights
OpenAI Guidelines: Building AI-Native Engineering Teams
The OpenAI official guidelines indicate that software development has entered an era of 'agent-led execution, with humans responsible for review and decision-making.' It states that approximately 80% of the software development lifecycle consists of repetitive tasks that can and should be handled by coding agents. The guidelines emphasize that the value of engineers is rapidly shifting from 'writing code' to defining problems, reviewing, and making decisions, requiring teams to adapt their organization and processes to accommodate agent-led workflows.
Cursor 2.1: AI Review and Instant Grep Search in the Editor
Cursor 2.1 has been released, focusing on AI code review within the editor, local instant Grep search, and improved planning mode. AI code review can be run directly on the current file or selected changes, automatically identifying potential bugs, performance issues, readability suggestions, and security risks, displayed in a side panel. The planning mode actively raises clarification questions when generating modifications for multiple files or complex refactoring plans, and supports full-text content location after plan generation via ⌘+F. Instant Grep search has been changed to real-time execution, supporting regex and word boundaries, with faster search and AI interaction responses, making version 2.1 more focused on functional enhancements.
Gemini 3 Agent Building Practical Guide (Phil Schmid)
Phil Schmid has released a practical tutorial on building agents from scratch using Gemini 3, aiming to demystify the agent-building process. Schmid emphasizes that the essence of an agent is 'an LLM placed in a loop and connected to several tools.' By breaking down the implementation steps and tool calls, he aims to make the agent-building process systematic and reusable, lowering the entry barrier.
Vibe Coding Guide: Planning Ahead to Avoid Code Chaos
The Vibe Coding practical guide on GitHub points out that while using AI to develop complex projects from scratch may go smoothly initially, the code can become chaotic and difficult to maintain after iterations. The core of the guide is 'planning ahead,' managing project context through design documents, technology stack selection, implementation plans, and the Memory Bank mechanism. It suggests breaking large projects into testable small steps and provides debugging and fixing tips, along with usage recommendations for tools like Claude Code and Codex CLI.