Managing Multi-Project Agent Workflows in VS Code

The Agent Window Interface

The new Agent Window in VS Code provides a centralized control center for managing multiple agent sessions across different workspaces, remote repositories, and local projects. It addresses the cognitive overhead of context switching by allowing developers to monitor the status of various agents—such as PR reviews or custom instruction tasks—within a single, unified view. The interface supports sorting by update time or workspace and integrates with existing VS Code themes, settings, and extensibility hooks.

Token Optimization and Model Routing

To improve cost efficiency, the VS Code team implemented automatic model routing within the 'Auto' model setting. This system analyzes task complexity and dynamically routes requests to the most appropriate model, resulting in a roughly 10% token cost reduction compared to using larger models for every step. The recommended workflow is to use a high-capability model for initial planning and then delegate implementation tasks to the Auto model. The team is currently exploring ways to surface these savings to users, similar to a retail receipt, to provide visibility into the cost-efficiency of their agent workflows.

Custom Hooks and Notifications

Developers can extend agent functionality using custom hooks to manage notification fatigue. One specific implementation involves using a local speech model hook that triggers upon task completion. This hook calls a task-completion function with a summary of the work performed, which is then processed by a local speech model to provide an audio notification of what the agent specifically accomplished. This allows users to track agent progress without relying on generic system sounds.