# Cooking with Agents in VS Code Liam Hampton, Microsoft | AI Engineer Europe 2026 | 17:04 Source: https://www.youtube.com/watch?v=dyHpnnlkTc8 Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/cooking-with-agents-in-vs-code Published: 2026-05-21 Tags: agents, coding-agents, human-in-the-loop, mcp ## TL;DR - Local agents fit work that needs close human review, such as writing and updating tests. - Background agents use isolated Git worktrees for larger tasks where the developer wants partial involvement. - Cloud agents run in GitHub Actions and fit work such as documentation when the developer does not need to stay involved. ## Summary Liam Hampton presents VS Code as one place to work with local, background, cloud, and third-party agents. He assigns each path according to how much involvement he wants. A local agent running Claude Opus writes tests while he reviews and redirects it. A background agent uses a Git worktree to build a front end from a GitHub issue, with Hampton checking the result before a pull request. A cloud agent runs in GitHub Actions to add open-source files without requiring his attention. The demo shows all three working on one Python product-store application at the same time. Hampton also explains the controls around cloud agents, including isolated environments, restricted network access, no direct access to the main branch, and MCP connections for GitHub and Playwright. VS Code brings agents, custom instructions, skills, hooks, prompt files, and MCP servers into one control surface, with support for Claude and other services alongside Copilot. ## Key ideas ### Agents still need a developer in the loop [00:55](https://www.youtube.com/watch?v=dyHpnnlkTc8&t=55s) Hampton rejects the idea that an agent can solve an entire problem from one prompt. Developers still need to think about context, tool access, flexibility, and token spend. He connects this to business questions about return on investment and productivity. His point is practical: teams need to choose how they use these systems instead of assuming that more AI infrastructure automatically produces value. The talk focuses on GitHub Copilot agents, but he says the same ideas apply to other agent services. ### The amount of human involvement should determine the agent path [02:24](https://www.youtube.com/watch?v=dyHpnnlkTc8&t=144s) Hampton divides agent work into three paths. A local agent runs in VS Code and keeps the developer close to the code, which suits writing tests. A background agent works in an isolated Git worktree, allowing a developer to stay partly involved while it handles a larger task such as building a front end. A cloud agent is for work the developer does not want to touch, such as documentation or making a repository open-source friendly. The choice depends on how much oversight the task needs. ### A background agent can build a front end from an issue [05:45](https://www.youtube.com/watch?v=dyHpnnlkTc8&t=345s) Hampton starts with a simple Python CRUD product store and asks a background agent to summarize and plan a solution to GitHub issue 25, which requests a front end. The agent works through the GitHub Copilot CLI inside VS Code and uses Autopilot to avoid repeated approval prompts for tool calls. Hampton tells it to pause before creating a pull request so he can test locally. The work happens in a Git worktree, a branch mapped to an isolated folder. ### The three agents can work on separate problems at the same time [07:18](https://www.youtube.com/watch?v=dyHpnnlkTc8&t=438s) While the background agent builds the front end, Hampton starts a cloud agent to add a README, contribution guidelines, and other files needed for an open-source project. He then starts a local agent with a custom Python testing agent and Claude Opus 4.6 at medium reasoning. The local agent writes unit tests, runs them, and updates route error handling when Hampton asks. Hampton can review that work while the other two agents continue, all from VS Code. ### Cloud agents are isolated GitHub Actions work [11:43](https://www.youtube.com/watch?v=dyHpnnlkTc8&t=703s) Hampton explains that cloud agents run in GitHub Actions inside an isolated environment. They can use the GitHub MCP server and Playwright MCP server, which supports tasks such as front-end testing with screenshots. Network access is restricted with firewalls and allow-listing. The agent cannot access the main branch, so it cannot push directly there. These controls let Hampton hand off documentation and similar tasks without giving the agent unrestricted access to the repository. ### Agent configuration is built from reusable pieces [12:56](https://www.youtube.com/watch?v=dyHpnnlkTc8&t=776s) The workflow can be shaped with custom instructions, custom agents, prompt files, and agent skills. Hampton describes custom agents as focused configurations for tasks such as writing or fixing test cases. Agent skills are presented as a newer form of the guidance people previously put in agents.md. VS Code also exposes hooks and MCP servers. These settings let developers define how agents work instead of relying on a blank chat prompt each time. ### VS Code puts different agent services behind one control surface [13:30](https://www.youtube.com/watch?v=dyHpnnlkTc8&t=810s) Inside the VS Code chat settings, Hampton shows built-in agents such as ask, explore, and plan, along with his custom test agent, skills, extensions, instructions, prompts, hooks, and MCP servers. The same area can include third-party support. He shows Claude plugins, hooks, instructions, and skills alongside Copilot. His claim is that developers can manage these different entry points from one VS Code interface rather than switching between unrelated tools. ### Awesome Copilot provides examples that can be adapted [14:57](https://www.youtube.com/watch?v=dyHpnnlkTc8&t=897s) Hampton points to the open-source Awesome Copilot project at aka.ms/awesomecopilot. Although it is aimed at Copilot, he says its configurations can be adapted for other AI tools. The project is also available through an MCP server. He explains that MCP can connect models and chat customizations to resources such as Azure, GCP, AWS, Playwright, documentation systems, and Microsoft Learn, with authentication or open access depending on the server. ## Notable quotes - "Absolutely not the case." (00:55) - "So what I'm really looking at is VS Code as a single entry point for AI agents." (04:49) - "One code base, three problems, three separate agents fixed all at the same time." (11:02) - "They also doesn't have access to your main branch." (12:17) ## Tools & references mentioned - GitHub Copilot - VS Code - Claude Opus 4.6 - GitHub Copilot CLI - Autopilot - GitHub Actions - Git worktrees - MCP - GitHub MCP server - Playwright MCP server - Awesome Copilot - Microsoft Learn - agents.md ## Who should watch - You are deciding whether a task belongs with a local, background, or cloud agent and want a concrete division of responsibility. - You work in VS Code and want Copilot, Claude, custom agents, skills, hooks, and MCP servers in one interface. - You need to hand off documentation or repository maintenance while keeping code changes isolated from the main branch. ## Related talks - [Collaborating with Agents in Your Software Development Workflow](https://aietalks.com/talks/collaborating-with-agents-in-your-software-development-workflow) (Jon Peck & Christopher Harrison, Microsoft, 1:04:06) - [The Agent Awakens: Collaborative Development with Copilot](https://aietalks.com/talks/the-agent-awakens-collaborative-development-with-copilot) (Christopher Harrison & Jon Peck, GitHub, 1:04:06) - [Containing Agent Chaos](https://aietalks.com/talks/containing-agent-chaos) (Solomon Hykes, Dagger, 23:48) - [Real World Development with GitHub Copilot and VS Code](https://aietalks.com/talks/real-world-development-with-github-copilot-and-vs-code) (Harald Kirschner & Christopher Harrison, Microsoft, 1:19:33) - [Agentic Engineering: Working With AI, Not Just Using It](https://aietalks.com/talks/agentic-engineering-working-with-ai-not-just-using-it) (Brendan O'Leary, Kilo Code, 27:03)