Cursor 3 subagents and skills

Subagents can separate a bounded investigation from the parent task. Skills can package a repeatable procedure. Neither feature guarantees faster work or a reviewable result. The useful question is whether the parent can verify the child’s output without reconstructing every intermediate step.

Start with one evidence-producing task

Choose an investigation with a clear answer, such as finding all callers of a deprecated API. Ask the child to return file paths, function names and any uncertain matches. Keep the first trial read-only. Compare its list with an independent search and inspect the disputed cases.

If the child edits code in a later trial, require the changed paths, actual verification output and remaining uncertainty. Read the combined diff. Two individually plausible changes can still disagree about an interface or test fixture.

Resolve rule conflicts where they occur

Inspect applicable project rules, scoped rules and the skill instructions. Remove contradictory guidance or narrow its scope. Do not claim a sentence at the top of a skill changes the client’s instruction hierarchy. A rule also cannot revoke a tool’s actual credential or filesystem access.

Keep repository-wide rules for conventions that apply everywhere. Scope local conventions to the files they govern. A broad glob is not automatically wrong; the question is whether its instruction is relevant and consistent in every matched location.

Make connector access reviewable

For an integration used in the task, record its owner, intended resources and operations, and how access can be revoked. Check the actual account and token permissions. Test an out-of-scope operation against synthetic data before relying on a claimed read-only boundary.

A temporary exception needs a removal condition and someone responsible for verifying removal. Merely writing a retirement date in a comment does not disable access on that date.

Evaluate the first handoff

Give a reviewer the task, the child’s report and the actual diff or search evidence. Ask which conclusion they can verify and which one still requires the original operator. Fix that missing evidence before increasing concurrency or adding another skill.

Use the Cursor agent documentation, rules reference, skills reference and MCP guide for client mechanics. The subagents and skills topic connects these checks to a team workflow.

Updated 21 September 2026: removed unsupported speed guarantees and the misleading skill-precedence template; added a concrete search and handoff verification exercise.