Use an AI coding agent in VS Code: your first Andromity task

An AI coding agent can inspect a repository, propose changes, edit files, and run commands. Andromity puts that workflow in VS Code’s secondary sidebar so you can keep the editor and file explorer visible. This guide starts with one small change you can review and test.

1. Install the VS Code extension

Find Andromity AI Coding Agent in the Visual Studio Marketplace, or install it from the command line:

code --install-extension agenticmarket.andromity-agent

Open a repository you own. Open the Andromity sidebar and follow its runtime setup prompts. The installation guide covers the terminal runtime and supported platforms.

2. Choose the model and provider

Use the model picker to select your provider and model. If you bring your own API key, enter it through the provider settings rather than pasting it into a task prompt. Cloud usage follows your provider’s pricing and data policies. For local inference, follow the Ollama coding agent guide.

Start with a model that supports tool calling. Model choice affects how reliably the agent edits code and follows the task. Andromity supplies the repository tools and approval workflow; the model supplies the decisions about using them.

3. Trust the workspace and keep SAFE mode

Grant trust only after you have reviewed the repository. Folder trust is checked before execution permissions: an untrusted folder blocks repository tool reads, writes, and command execution. Changing the permission mode does not bypass that boundary.

Use SAFE mode for the first task so you can review actions as they are requested. See workspace trust and permission modes for the complete security model.

4. Give the agent a testable task

Prefer a bug with a reproducible input over a request to improve the entire project. Include the expected behavior, a narrow scope, and a test command if you know it:

The search box crashes when the query is empty. Find the cause, propose a minimal fix, and add a regression test. Keep the existing search behavior for non-empty queries. Run the relevant test and report its output.

Review the proposed plan before the agent edits. Give feedback if the plan changes unrelated files or adds dependencies unnecessarily. A useful plan explains which files it expects to change and how it will verify the result.

5. Follow execution and review the diff

The live execution trace helps you see tool calls and their timing. Use it to distinguish a model waiting on a command from a task that is still producing a response. The native diff view lets you inspect the actual edits alongside the original code.

Before accepting a change, check the following:

• The diff addresses the bug and stays within the requested scope.
• The regression test fails for the original problem and passes with the fix.
• The command output shows which checks ran and which were skipped or failed.
• No generated file, credential, or unrelated formatting change slipped into the patch.

If the approach is wrong, ask for a correction or use the available turn rollback when a checkpoint exists. Review the working tree afterward; rollback is not a substitute for version control.

Autocomplete, chat, or agent?

Autocomplete suggests code as you type. Chat explains code or answers a question. An agent coordinates multiple steps using tools, such as reading a function, editing it, and running a test. Use the smallest workflow that fits the work: a short question may need chat, while a reproducible bug benefits from the agent’s edit-and-test loop.

Where to go after the first task

Keep the next task similarly bounded. Once the workflow is familiar, explore MCP tool integrations, provider configuration, and terminal commands.