Microsoft is celebrating its 50th anniversary by launching Agent Mode with MCP support for all VS Code users worldwide. Our experts at ELEKS have worked with the tool for some time and tested its capabilities across multiple projects.
As it becomes available to the broader development community, we've asked Ostap Elyashevskyy, ELEKS' Test Automation Competence Manager, to share the insights.
How effective has Copilot been at catching bugs or improving your code?
GitHub Copilot’s code review agent is moderately effective at catching bugs and suggesting improvements, especially when it is used alongside traditional review processes. It can be helpful in identifying common anti-patterns, unused variables, and suggesting more readable alternatives to existing code.
The strength of the GitHub Copilot review
1. Syntax check. It can suggest more readable constructions and check the typos in variables, functions, classes, and other entity names.
2. Logic suggestions. A review agent can detect and notify about missing checks, inconsistent logic, especially in simple functions.
Let's look at the first example. GitHub review recommends checking the specific value instead of using general verification. Without this change, this test does not bring much value.
The Copilot code review proposed only error handling, as was shown above in the Copilot suggestion screenshot. But if we “Ask copilot” with the next prompt: “what else can be improved in this code ?”, the result will be more comprehensive and will propose changes by categories: Error Handling, Input Validation, Security, Response Data, Logging, Asynchronous Function Optimization, Use Constants for Status Codes, Documentation and Unit tests. All these suggestions are reasonable and make sense (for more details about recommended code changes, see screenshots attached).
Does Copilot offer deep code review insights or just surface-level feedback?
GitHub Copilot (as of now) tends to focus more on surface-level suggestions, like syntax, naming, formatting, and common best practices, rather than deep code review feedback. It cannot substitute human code review. GitHub review is good at suggesting better names for variables and functions, pointing to unused imports, offering minor changes, and recommending well-known best practices. GitHub review has difficulties with business logic, complex security and performance issues, architecture design, and some types of testing, such as edge-case testing.
What's your experience with managing PRs after AI reviews?
After the AI review, we still need to review the code and suggestions generated by the AI. It does not mean that AI is not helpful, but at the same time, we cannot rely only on AI. It can suggest chunks of code that work in isolation but don’t fit well together, so you need to modify and merge it. AI doesn’t always see the bigger picture, sometimes we have business constraints or legacy systems with their own coding approaches. Sometimes you spend more time reverse-engineering the AI's logic than writing it from scratch.
Related insights
