Agentic Development in VS Code: Live Demo and Workflow
Visual Studio Codego watch the original →
the gist
A live demonstration of the new agentic development features in VS Code, focusing on the 'Agents Window' for multi-session management, local agent hosting, and remote connectivity via dev tunnels.
The Shift to Agent-First Development
The session highlights the evolution of GitHub Copilot within VS Code, moving beyond simple chat interfaces toward an 'agentic' workflow. The team introduces the new 'Agents Window,' a dedicated workspace designed to manage multiple AI sessions and projects simultaneously without cluttering the primary editor. This interface allows developers to maintain high-level context across different tasks, effectively treating the AI as a partner that can be supervised rather than just a code-completion tool.
Managing Context and Discovery
A core theme is the move away from manual prompting toward agent-led discovery. Instead of laboriously describing a UI or a codebase, developers can point the agent to a live URL or a reference implementation. The agent then parses the DOM or codebase to understand the design patterns and requirements. This approach is not only more efficient for the developer but also more token-efficient, as it avoids the overhead of sending multiple high-resolution screenshots to the model.
Local vs. Remote Agent Execution
The speakers demonstrate the flexibility of the agentic architecture, specifically the 'Local Agent Host.' By decoupling the agent process from the VS Code UI, developers can run agentic tasks in a separate process on their machine. This setup supports advanced connectivity features like dev tunnels, which allow developers to access their local agent sessions from a browser via vscode.dev. This enables a seamless transition between local development and remote management, allowing users to monitor or trigger agent tasks from any device.
Balancing Control and Automation
The demo explores the spectrum of agent autonomy, ranging from 'manual approval' (where the agent requests permission for every tool call) to 'autopilot' (where the agent makes decisions to reach a goal). The speakers emphasize that the level of trust and automation should scale with the risk of the task—low-stakes projects allow for higher autonomy, while critical codebase changes require more rigorous verification loops.