VS Code Insiders Podcast Subagents: Parallel Execution and Context Isolation [Updated]
29 snips
Feb 9, 2026 
Harald Kirschner, a VS Code engineer focused on agent workflows and AI integrations, explains subagents and context isolation. He breaks down agent loops and context windows. He describes how subagents run in isolation, summarize findings, and execute in parallel. He also compares custom agents and model choices for different tasks.
AI Snips
Chapters
Transcript
Episode notes
How The Agentic Loop Builds Context
- The agentic loop assembles a system prompt, user question, tool calls, and responses into a growing context window.
- That accumulated context drives reasoning and determines what the agent can do next.
Tool Calls Expand Context Dynamically
- Tool calls (search, read files) pause the loop and inject new information into the conversation.
- Parallel tool calls let the agent scale by returning multiple results before continuing reasoning.
Why Context Isolation Matters
- Not all gathered context belongs in the main agent; much of it is intermediate and bulky.
- Subagents were introduced to isolate that bulk research and return only concise, relevant summaries.