Cursor 3 subagents and skills

Turn on Cursor 3 subagents and skills and your agents get faster. Whether your team can explain what they did is a separate question, and that gap is the thing to manage. Cursor is Anysphere's AI code editor. Subagents are scoped agent runs that handle a slice of work on their own, and skills are the SKILL.md playbooks that tell those runs how your team actually works.

Here is the trap. A green build means the code passed, not that a reviewer can trace why the agent edited a deploy script or which .mdc rule governed the change. The transcript that held the answer scrolls away. By the time someone asks, it is gone.

So the real work is not picking better prompts. It is writing four small contracts that put the answer in a file instead of in a chat log. Each one is a few lines. Each one pays for itself the first time someone has to defend a merge they did not write.

Name a steward for every connector

Where Cursor MCP is on, connectors multiply faster than ownership. Hosts approve new capability in a click; teams assign owners much slower. The gap fills with anonymous convenience.

Close it with one line per connector: a human owner and a rotation schedule, written next to the MCP entry. Use the Model Context Protocol specification as the shared boundary language so everyone means the same thing by "scope." New connectors that arrive without a steward line do not land.

Put broad rule globs on a diet

When .mdc globs are wide, agents inherit guidance that contradicts itself. Everything is active, so nothing is trusted, and a reviewer cannot say which file is in charge.

Split rules by folder concern. Ban the mega-glob unless it ships with explicit exceptions. The payoff is review speed: a reviewer can cite the governing file by name instead of running a scavenger hunt across every rule you own. Most of what feels like slow review is really just that hunt.

Give demo prompts an expiry date

Demos ship fast, and the temporary prompt that made the demo work quietly becomes a default. Shortcuts calcify when nothing tells them to retire.

Prefix any risky prompt with an owner and a retire-by date, written inside the rule comment. Now cleanup is scheduled instead of heroic, and the prompt explains its own shelf life to the next person who reads it.

Open every skill with a precedence preamble

Anyone using Cursor Skills has watched a skill contradict an always-on rule. Two sources of truth fire at once, with no ritual for deciding which wins.

Fix it mechanically: every SKILL.md opens by naming the .mdc files that override it. Precedence gets read from the file instead of remembered by whoever wrote it. Here is a small delegation snapshot you can adapt to your repo:

---
description: Delegation boundary snapshot (adapt globs to your repo)
globs:
  - "**/*"
alwaysApply: false
---

- Cursor: keep scopes explicit in `.mdc`; forbid undeclared MCP domains.
- Claude Code: cite `CLAUDE.md` precedence before expanding bash scope.
- Codex: ensure `AGENTS.md` carries replay-friendly verification notes for CLI runs.

The handoff-side version of these fixes lives in Cursor subagents and skills for teams, and the wider patterns sit in the subagents and skills topic hub.

Run the review gates

The agent surface is documented under Cursor Agent, org defaults under Cursor Enterprise, and surface changes in the changelog. The gates, though, are yours to enforce:

Before you merge agent work, walk this list:

• Forked agent work lists parent and child responsibilities.
• Red-folder paths got explicit human acknowledgement.
• Scopes in the PR body match the folders in the diff.
• Primary-doc links were smoke-checked after publishing edits.

None of this replaces architecture judgement. Agents speed up execution; ownership stays with you. Picture two clocks on the wall, one for shipping and one for explainability. When only the shipping clock ticks, the difference accrues as debt, and you pay the interest in review hours.

Common questions

• What do Cursor 3 subagents and skills change for a team?

  They move the hard question from prompting to governance. Subagents are scoped agent runs and skills are repo-owned playbooks, so the work becomes deciding who stewards each connector, which .mdc file governs each folder, when temporary prompts expire, and which rules override which skills. The tooling ships. The contract is yours to write.

• Why do skills and rules contradict each other?

  Because different people write them in different rooms, and the files inherit the gaps between those conversations. The fix is mechanical, not social: every SKILL.md opens with a preamble naming the .mdc files that override it. Precedence then gets read from the file instead of remembered by whoever happened to author it last.

• How do you stop MCP sprawl?

  Put a steward and a rotation schedule next to every connector entry, and refuse any connector that arrives without one. Hosts approve new capability faster than teams assign owners, and that speed gap is where sprawl grows. The steward line closes the gap and turns anonymous convenience back into infrastructure someone owns.

• What should a rollback rehearsal check for agent work?

  Whether the team can answer, from artifacts alone, which rule governed the change, which commands prove the regression guards, and who approved any red-folder paths. If those answers live only in transcripts that have scrolled away, the rehearsal failed before the rollback even started. Receipts in files are the whole point.

Where to go next

Install one fix this week. The skill preamble is the cheapest, so start there, then re-run your rollback rehearsal with receipts on. Our training runs that rehearsal against your own repo.