I have been defaulting to using GitHub Copilot for reviewing PRs, after all, GitHub runs it automatically and it doesn’t seem to cost anything.

But in the last few reviews, I started to doubt what it actually “understands.”

For example, a very basic issue: it still considers 1.82.0 to be higher than 1.91.1, which is a common version number comparison error from early large models.

If this were a model problem, it would also think that rust 1.91.1 has not been released yet, which reveals that the agent’s retrieval and real-time status judgment capabilities are also lacking.

Another bigger issue is: Copilot’s review is clearly based on individual files.

It’s okay for checking code style and boundary conditions, but it lacks a global perspective. For instance, in one PR, the agent miscalculated relative paths and copied the same file multiple times, when in fact only one copy is effective — it completely failed to notice this, and didn’t even care about what the original issue associated with the PR was asking for.

In my opinion, a qualified code reviewer agent should primarily evaluate from a global perspective:

Does the PR satisfy the issue, align with the project goals, and have a reasonable file structure and architecture? Only then should syntax and detail issues be considered.

Recently, I’m planning to add a reviewer mode to Holon.
Are you all really using reviewer agents now? What do you usually use?