Cooking with Agents in VS Code — Liam Hampton, Microsoft

May 21, 202617:046,962 views

Summary

Liam Hampton, a developer advocate at Microsoft, presents a framework for choosing between three categories of AI agents in VS Code — local, background, and cloud — and demonstrates all three running simultaneously against a single Python CRUD codebase. The session opens by challenging the assumption that agents can solve problems in one-shot prompts, and frames the real challenge as matching agent type to task context to get genuine ROI on AI infrastructure spend.

Hampton defines the three agent categories: local agents run on the developer's machine with a remote model, are fully interactive, and keep the human in the loop at every step; background agents use Git worktrees to operate in an isolated directory-mapped branch and allow a 50/50 hands-on/hands-off dynamic; cloud agents run in GitHub Actions inside a sandboxed, network-firewalled environment with no access to the main branch, and are appropriate when the developer genuinely does not care about the implementation path.

The live demo centers on a simple Python product store with no frontend, no tests, and no documentation. Hampton triggers three agents simultaneously: (1) a local agent using Claude Opus 4.6 with medium reasoning to write and iterate on unit tests, giving hands-on to-and-fro with error handling fixes; (2) a CLI background agent running in Autopilot mode (preview feature that suppresses tool-call confirmation prompts) against a Git worktree to build a new React frontend from a GitHub issue, instructed to pause before opening a pull request so Hampton can test locally; and (3) a cloud agent tasked with making the repository open-source friendly by generating a README, contribution guidelines, and related files. The cloud agent runs in GitHub Actions, has access to the GitHub MCP server and the Playwright MCP server for automated frontend testing with screenshots, is whitelisted via network firewall, and cannot push directly to main.

Hampton emphasizes VS Code as a single control-plane entry point that surfaces local, background, remote, and third-party agents — including non-Microsoft models and tools — through a unified chat pane and settings modal. The settings modal exposes custom agents (scoped instruction sets for specific tasks like test authoring), prompt files, agent skills (described as the successor to agents.md), hooks, and MCP server connections. Hampton shows a custom test-writing agent pre-loaded for the Python project and demonstrates how Claude and other third-party providers are accessible from the same configuration surface alongside GitHub Copilot's built-in ask, explore, and plan agents.

Key practical points raised: Autopilot mode is powerful but should be used cautiously due to unrestricted tool-call execution; background agents using Git worktrees cleanly isolate experimental work without contaminating the main working tree; cloud agents in GitHub Actions are safe for open-ended tasks because of their sandboxing and branch restrictions; and the MCP ecosystem (GitHub MCP, Playwright MCP, Azure, GCP, AWS integrations) significantly extends what cloud and local agents can interact with. Hampton also mentions the open-source Awesome Copilot project (aka.ms/awesomecopilot) and a companion MCP server as resources for customizing agent behavior across any AI tooling, not just GitHub Copilot.

Key Takeaways

•Match agent type to task: use local agents when you want human-in-the-loop iteration (e.g., writing tests), background agents for large isolated tasks you want to monitor but not drive (e.g., building a frontend), and cloud agents when you want full delegation (e.g., documentation).
•Git worktrees enable background agents to work in a completely isolated directory-mapped branch, preventing contamination of your main working tree while you continue developing in parallel.
•Cloud agents in GitHub Actions provide built-in safety through network whitelisting, branch protection (no direct push to main), and sandbox isolation — making them appropriate even for open-ended tasks.
•VS Code's settings modal provides a single control plane for configuring local, background, cloud, and third-party agents, including custom agents, prompt files, skills, hooks, and MCP servers.
•Autopilot mode in GitHub Copilot CLI suppresses tool-call confirmation prompts to speed up background agent work — use it deliberately and understand the risk of unrestricted tool execution.
•The GitHub MCP server and Playwright MCP server are available to cloud agents, enabling automated frontend testing with screenshots directly from a GitHub Actions workflow.
•Custom agents (scoped instruction sets) let you pre-configure task-specific expertise — such as a Python test-writing agent — that can be invoked directly from the VS Code chat pane.
•Running three agents simultaneously on one codebase is practical today in VS Code: one writing tests, one building a UI from a GitHub issue, one generating documentation — all visible from one interface.
•Third-party models including Claude are accessible from the same VS Code agent configuration surface as GitHub Copilot, so the local/background/cloud framework applies regardless of which LLM provider you use.

Transcript

Hello everybody. It's great to see so many of you who are still here at the final on final day right at the end. So I hope you all had a great conference. Show of hands, who here uses GitHub Copilot? Awesome. Lovely stuff. Who here uses VS Code with GitHub Copilot? Awesome. So I'm going to be talking about both these things today. I'm going to be talking about cooking with agents in VS Code. Now gentleman before was speaking about cognitive load of agents and that is absolutely correct. You see ...