I'm currently building one based on: https://blog.cloudflare.com/ai-code-review/

The keys here are

1. Use mechanical scripts at the start to gather a context and at the end to make the comments. Don't have an agent/llm do this part (make API calls), I do have the comment creation agent write a jsonl file, so it is writing the comments.

2. It should not need any tools beyond read and write (to a specific directory). Tests and the like should have already run as part of CI. PR review is not really intended review failing tests, human or agentic.

3. A way to run different level of agentic review based on how complex the PR is.

General PR agents are not really the right answer in my opinion, in large part because devs and orgs are opinionated and those need to be encoded into their PR review process.