← back
Cooking with Agents in VS Code — Liam Hampton, Microsoft
Takeaway
Choose agent execution modes according to how much supervision each task needs and coordinate them from a shared development environment.
Summary
- VS Code provides a common entry point for local, background, and cloud coding agents, with different levels of developer involvement.
- Background Copilot CLI agents use Git worktrees for isolated changes; Hampton uses them for UI work while reserving local agents for closely supervised tests.
- A demo runs three workflows together: a background agent builds a frontend, a cloud agent adds open-source documentation, and a local agent improves tests and error handling.
- Autopilot reduces tool-call interruptions, but the example explicitly instructs the agent to pause for local testing before creating a pull request.
github-copilotvscodecoding-agents
Original description
One codebase, three problems, three agents running at the same time. Liam Hampton from Microsoft demos the full loop in VS Code: a local agent with Claude Opus writing and fixing unit tests with him in the loop, a background agent using a git work tree to build a front end from a GitHub issue without him touching it, and a cloud agent running in GitHub Actions to make the repo open source friendly. The talk is a framework for knowing which agent path to pick and why. Local when you want hands on iteration. Background when the task is big and you can tolerate being half in half out. Cloud when you genuinely do not care how it gets done. VS Code handles all three from one interface, with Copilot, Claude, and third party agents accessible from the same control plane. Speaker info: https://x.com/liamchampton / liam-conroy-hampton https://github.com/liamchampton