Inside The Agent Loop with Pierce Boggan

Visual Studio Code hosts a discussion where James and Pierce Boggan break down how VS Code’s “agent loop” works under the hood for AI-driven coding assistance, including how agents coordinate tools and sub-agents, and what you can tune in your own workflow.

Full summary based on transcript

Introduction and agent loop overview

The presenters introduce the concept of an “agent loop” in VS Code: an iterative process where an AI agent plans work, takes actions (often via tools), evaluates results, and repeats until it reaches a stopping point.

Tools, modes, and agent optimization

They discuss how tools and different operating modes fit into the loop, and how the loop has evolved over time.

Key points covered:

Sub-agents and harnesses

They explain how sub-agents can be used to break work into smaller, specialized tasks, and how “harnesses” help coordinate or structure how agents and tools are used together.

Why use different models for sub-agents?

They cover the rationale for selecting different models for different tasks in the loop.

Topics mentioned:

Customization, trade-offs, and closing thoughts

They close by discussing how developers can customize or optimize their workflows, and the practical trade-offs that come with those choices.

Video chapters (from the description):