Cursor Subagents and Skills for Engineering Teams

We noticed in Cursor workshops that teams rarely fail because a subagent cannot produce code. They fail because nobody can explain the delegation after the diff exists. Cursor subagents and skills are a delegation system, so the hard part is making the work reviewable after the agent is done.

Operating Model

Use this review protocol with three layers:

The layers should not compete. Rules set the floor, skills explain task behavior, and subagents execute within a named scope.

Scope Ledger

Every parallel run needs a scope ledger before the first subagent starts.

# Cursor subagent scope ledger

- Parent intent: update onboarding analytics without changing auth behavior.
- Subagent A: UI copy in `apps/site/src/app/onboarding/**`
- Subagent B: analytics tests in `tests/onboarding/**`
- Forbidden paths: `apps/site/src/lib/auth/**`, `.env*`
- Relevant rules: `.cursor/rules/analytics.mdc`
- Relevant skill: `skills/event-review/SKILL.md`
- Proof command: `pnpm test onboarding`
- Merge owner: product engineering lead

This is the smallest artifact that turns parallel work into a reviewable workflow.

Rule Precedence

State precedence plainly:

1. Human safety decision.
2. Repo rules.
3. Skill instructions.
4. Subagent task prompt.

If a skill contradicts a repo rule, the rule wins. If a human reviewer identifies risk, the reviewer wins. Without this order, teams end up debugging which instruction the agent believed.

Review Evidence

The reviewer should be able to inspect four things without replaying chat:

• the scope ledger;
• the rule or skill that applied;
• the diff;
• the verification command and result.

This connects directly to our methodology: Review should reconstruct intent from artifacts, not from memory. Teams building a broader operating model can pair it with the subagents and skills topic hub.

Training Exercise

Run this as a team exercise:

Exercise: bounded agent delegation

1. Choose a low-risk module.
2. Write one `.mdc` rule for that module.
3. Write one skill for a repeated task.
4. Launch one subagent with a path limit.
5. Paste the scope ledger into the PR.
6. Ask a reviewer who did not see the chat to explain the diff.

If the reviewer cannot explain the work, the system is not ready for higher-risk tasks.

Sources

• Cursor Agent
• Cursor Rules
• Cursor Skills
• Cursor MCP
• Cursor changelog
• Model Context Protocol specification

Best ways to use this research

• Best for: engineering teams comparing Cursor, Claude Code, and Codex operating habits under delivery pressure around “Cursor Subagents and Skills for Engineering Teams.”
• Best first artifact: turn the named fix into a shared checklist, repo rule, handoff receipt, or policy table before the next automated run.
• Best comparison angle: compare the workflow across review evidence, connector scope, and handoff friction before adding another agent run; keep the path that leaves the shortest auditable trail.

Next Step

Start with the subagents and skills topic and run the training exercise against one real branch. Keep the first run small enough that the team can inspect every artifact.