Plan & Act: How Devs Work with AI

Origins of Plan & Act Modes 00:00

• The concept of having two distinct modes for developer interaction—plan and act—was first introduced by Klein.
• Users naturally began using the AI agent by creating a markdown file to outline architecture or plans before starting work.
• This emergent workflow inspired the team to build more intuitive support for such modes into the product to help new users adopt the pattern easily.

How Plan & Act Modes Work 00:36

• Plan mode is designed for exploration, where the agent gathers information, reads files, and builds context to formulate a plan for the user’s task.
• Act mode directs the agent to follow the established plan, executing commands and making edits as needed.
• Users often primarily engage with the agent in plan mode, with significant back-and-forth for context extraction.
• During plan mode, the agent may ask clarifying questions to fully understand requirements, such as project themes or desired website pages.

User Experience and Workflow 01:39

• Once the user is satisfied with the plan, they switch to act mode, can enable auto-approve, and allow the agent to execute the work autonomously.
• Most user involvement occurs during the planning phase; during act mode, users mainly monitor or intervene if corrections are needed.
• This workflow lets users delegate routine work and only step in when course correction is necessary, enabling a smoother collaboration with the AI agent.