Writing
Reviewing code an agent wrote.
What we have learned building an analysis engine for agent-written pull requests. Every claim here is one the engine can back, or it says plainly that it cannot.
How to review a pull request your AI agent wrote
A method for reviewing agent-written code: start at the system, not the diff. What to read first, what to skip, and how to know when you are done.
Your agent changed 60 files. Where do you start?
The diff is the wrong entry point when an agent wrote it. Start with the components that changed, then the ones they reach. Here is the order, and why.
Why LLM-drawn architecture diagrams are wrong
Ask a model to draw your architecture and it draws something plausible. The problem is not that it hallucinates — it is that you cannot tell where.
What leaves your machine when a tool calls an LLM
Most code tools that call a model are vague about what they upload. Here is the question worth asking, and the exact, checkable answer for one CLI.
Claude Code vs Cursor vs Codex: reviewing the PRs
The three agents produce differently shaped changes with different failure modes. What to look for in a pull request from each, and what does not change.
Concepts
The four ideas the posts above keep leaning on, each defined in one page.