Key takeaways

• Review agents like Bugbot live inside the PR workflow: auto-trigger on updates, read existing comments, publish CI status, and use project rule files (.cursor/BUGBOT.md) and learned memory.
• Cursor's security agents prove Chapter 1's thesis: a 15-line prompt catches hundreds of bugs because the infrastructure underneath (MCP server, deployment, orchestration, cross-run state) does the heavy lifting.
• The core reliability law: an agent cannot mark its own homework. Pair the probabilistic LLM (the researcher) with an independent deterministic check (the peer reviewer); you need both.
• Keep a human in the loop for anything that matters; even Cursor's review agent does not auto-push fixes.
• The components an agent depends on (MCP servers, skills, templates) are a new agentic supply chain and a real attack surface.

Original sources: Cursor's Bugbot docs and Snyk's analysis "I Read Cursor's Security Agent Prompts, So You Don't Have To."

---