Cursor 3.3 context for subagents and skills

The fastest way to keep Cursor 3.3 from clogging your merge queue is to give every context source an owner. In Cursor, Anysphere's AI code editor, context is the set of rules, skills, and MCP connectors that are active when an agent run starts. When nobody owns those sources, reviewers can't tell which rule or connector shaped a diff, and the queue slows to a crawl. The version is not the issue. Unowned context is.

See what loads into context

Before you assign owners, you need to know what is in the room. An agent run pulls from your .mdc rules, your installed skills, and any MCP connectors the host has approved.

Each of those can change behavior silently. A broad glob can hand the agent guidance that contradicts another rule. A connector someone added during a demo can still be granting capability months later.

So the first question for any run is simple: which sources governed this diff, and can I point to each one? If you can't cite it, you can't review it.

Write a PR receipt reviewers can trust

Cursor Blame tells you who changed a line, but it can't tell you why when the PR body is empty. The fix is a review receipt: a short PR description that lists the parent intent, the child scopes, and the verification transcript.

This moves the cost of understanding a change back to the author, where it is cheapest to pay. A reviewer who is new to the area can trace the intent without replaying a chat log.

Here is a small .mdc rule you can drop in to keep delegation boundaries explicit across agents:

---
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.

Give every connector a steward

Connectors multiply faster than ownership once Cursor MCP is enabled, because a host can approve a capability in seconds while assigning a human owner takes a conversation. Left alone, connectors land as anonymous convenience.

Name a steward beside each MCP entry, plus a rotation schedule. Now a change to a connector has a person attached to it, and "who approved this network access" has an answer.

Do the same trimming on rules. Broad .mdc globs feed agents contradictory guidance, so put your rules on a glob diet: split them by folder concern and refuse mega-globs without a written exception. The governing file becomes citable in one click.

Put a retire-by date on demo prompts

Vendor-demo-heavy quarters leave temporary prompts behind, and those shortcuts calcify when nothing forces them out. A prompt that made sense for one live demo can quietly steer real runs for months.

Tag each temporary rule with an owner and a retire-by date inside the rule comment. Housekeeping becomes a scheduled task instead of an act of heroism by whoever notices the mess.

A quick set of gates keeps this honest at review time:

And a short strip you can paste into a PR template:

• 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.
• MCP connectors mentioned (if any) list owners.

Hard constraints still belong to humans. Threat models, customer promises, and blast-radius calls stay off autopilot.

Common questions

• What changes with Cursor 3.3 for subagents and skills?

  Treat Cursor 3.3 as a prompt to audit context, not a feature hunt. The durable work is naming an owner for every context source: review receipts in PR bodies, a steward beside each MCP connector, rules split by folder concern, and expiry tags on leftover demo prompts. The release matters less than the ownership habits it nudges you toward.

• What counts as context in Cursor?

  Context is the set of rules, skills, and MCP connectors active when an agent run starts. The review question is which of those governed a given diff. So each source needs to be citable on its own: the .mdc file by folder, the skill by precedence, and the connector by its named steward. If you can't cite it, you can't review it.

• How do you cut merge-queue fatigue from agent runs?

  Charge each stream its review bill upfront. A PR is ready when its receipt is ready, with parent intent, child scopes, and a verification transcript included. Fatigue is the global cost of many locally fast streams, and a receipt moves that cost back to the author, who can settle it for far less than a tired reviewer can.

• Where should I standardize first?

  Start where blame is hardest to trace today. If reviewers keep asking why a change happened, write review receipts first. If network access feels untracked, assign connector stewards first. Pick the one fix that leaves the shortest auditable trail, ship it as a real file, then move to the next.

Start with one tired queue

Pick a single merge queue that feels slow, turn one named fix into a real .mdc rule this week, and see who carries less. The drills for going further live on Cursor subagents and skills.