A 9-question check for Cursor agent setups before they run unattended

When a Cursor agent starts running longer tasks, or you wire it into hooks, MCP tools and background agents, the failures are rarely about the model being “dumb”. They’re usually missing guardrails: a loop with no step limit, a retry that fires forever, a tool that can delete or publish with nobody in the loop.

Here are the nine questions worth answering before you trust a setup on its own:

  1. Is every model pinned to an exact model ID, not “latest” or an alias?
  2. Does each tool or API call have a timeout?
  3. Are retries capped, with backoff?
  4. Is there a maximum number of steps or loops per run?
  5. Is there a spend limit or a budget alert?
  6. Do you know whether your provider stores prompts, and for how long?
  7. Do delete, pay, send and publish actions need a human OK?
  8. Is the agent limited to an explicit list of tools (for MCP, only the servers and tools it actually needs)?
  9. Is there a fallback if the main model or provider is down?

For Cursor specifically, 4, 7 and 8 come up most often. Keep risky shell commands off the auto-run allowlist, give each MCP server the narrowest token it can work with, and put “stop and ask before anything irreversible” in your project rules, so it applies to every chat and background agent and doesn’t depend on remembering it.

If you want to check a setup quickly, we built a free checker that runs entirely in your browser (nothing you paste is sent or saved): Agent Setup Checker · Stackyard. Paste a config or answer the nine questions, and it shows what’s missing. Disclosure: it’s ours, from Stackyard. The matching rule and hook snippets are in an MIT repo, linked from the checker’s results page.

Happy to hear which of these people here handle differently, especially around background agents.