Adding rules that describe the project functionality, architecture and folder structure has proved beneficial in reducing mistakes, adding context to tasks. (hint: you can ask the agent to generate these if they don't exist)
So far I've found one useful MCP https://browsertools.agentdesk.ai/installation that helps the agent gather info from the browser during tasks (e.g screenshots, console logs, network errors) to help debug.
Wrote about it recently (https://davidgasquez.com/cursor-agent-tricks/) although is now outdated! You can find the latest setup on my Dotfiles repository (https://github.com/davidgasquez/dotfiles/tree/main/cursor).