The point about LLMs reinforcing their own assumptions matches something I keep running into: when the same model writes the code and the tests, the tests tend to encode the implementation, so they pass by construction. What has helped more than a better model is separating the roles, e.g. tests written from the spec or ticket before the code exists, or a reviewer pass that never saw the implementation prompt. Does Mo get the PR description or spec as input, or does it only explore the running app?