What makes an AI code review tool genuinely useful for developers?

I've been building Mesrai, a multi-agent AI code review platform, and one challenge I've been thinking about is how to make AI code reviews genuinely useful in real-world development.

AI reviewers can identify potential bugs, but developers still struggle with false positives, missing codebase context, noisy suggestions, and findings that aren't actionable.

I'm curious about the experiences of other developers here:

1. What's your biggest frustration with AI-powered code reviews?
2. Do you find whole-codebase context more useful than reviewing only changed lines?
3. Would you prefer multiple specialist agents reviewing different aspects of a pull request, or one AI reviewer?
4. What would make you trust an AI review enough to act on its suggestions?

I'm building in this space with Mesrai and would love to learn from people working with AI agents, developer tooling, and code quality.

If you're interested, here's the project: https://mesrai.com/

I'd especially appreciate real examples of what works, what doesn't, and what's still missing from existing tools.