From Ask to Agent: The Evolution of GitHub Copilot in VS Code

Note: This blog is a personal notebook. I write it to record what I work on, for my own reference, not to publish it on social media or to promote myself. If a post is still useful to you, I am glad, but that is not why it exists.

This post was published on November 26, 2025. Software versions, product features, links, and tool availability may have changed since then. Always check official documentation for the latest information. Some screenshots show small details. Zoom in on the page to see them clearly.

Copilot in VS Code changed what it does, not only how well it does it. First it answered questions. Then it proposed edits. Now it works across the workspace toward a goal you describe.


Problem

Many teams started with Copilot when it gave only quick answers and code examples, and they still use it that way. The result is slower delivery and more manual work than the tool allows.


Why this matters

Each mode moves your responsibility: from writing, to reviewing, to directing. In Agent Mode, the AI reads files, edits across the workspace, proposes commands, reacts to errors, and continues until the task is done.

This makes your value more visible. You know the business rules and the edge cases, and the agent still needs your direction and judgement.


How the modes compare

Mode What it does Your role
Ask Mode Answers questions and explains Implementer
Edit Mode Proposes direct file changes Reviewer
Agent Mode Works across the workspace toward a goal Director

Planning a trip works the same way. A guidebook answers questions. A travel agent proposes each booking for your approval. A tour organizer plans the whole trip from your brief.

Copilot can change files in your repository, but it does not know your client’s unwritten rules unless you state them.


Demo steps

Prerequisites

  • VS Code with the AL Language extension
  • GitHub Copilot with Agent Mode enabled
  • An AL workspace with a codeunit that has real logic

1. Select a task with clear limits, for example:

Write a test codeunit for codeunit 50200 "Invoice Validation". Cover valid invoice, missing customer number, zero amount line, and closed-period posting date. Use the BC test library.

2. Let Copilot make the first version

Copilot reads the codeunit and proposes a test file.

3. Review before you accept

Check that the tests examine the real business rule, not only a structure that compiles.

4. Compile and run

Compile and test as usual. The failures feed the next iteration.

The lesson is how precise your instruction must be before the result is reliable.


Risks and limits

  • It can make large changes quickly.
  • It uses only the context it can see.
  • It does not know the client’s business intent.
  • It can write code that looks correct but is not.

More capable AI needs stronger review habits, not weaker ones. Keep secrets and client data out of the agent’s scope.


Next iteration

  • Start with tests, documentation, and controlled refactoring.
  • Build a small prompt library for repeated AL tasks.
  • Agree on review rules for AI output.
  • Connect the workflow to compile and test feedback.

Closing note

Copilot changed how the work is divided. It used to answer. Now it acts. The AI does the mechanical part, and you define the goal, the limits, and the meaning of correct.

← Back to AI