VS Code engineers Ben and Peng discuss the evolution of AI agent orchestration in VS Code over the past year. They explain three types of agents: local foreground (interactive, responsive), background (autonomous with file isolation), and cloud (fully remote). The team has unified agent management in the chat view, allowing developers to run multiple sessions in parallel. Key challenges include helping users understand when to use each agent type, improving code review workflows for AI-generated changes, and making projects 'AI-ready' with proper linting, testing, and documentation. The conversation emphasizes treating agent sessions like email—triaging, archiving, and responding as needed—while the team continues refining the UX based on dogfooding and user feedback.
Sort: