Skip to lesson
Exit
Cursor team workflows in a shared repo1 / 2

2 min lesson

What the top two rungs actually are

Use the lesson to respond to this: "You're deciding whether to keep a job in the foreground (local, watched) or push it to a background cloud agent. What's the test?" Keep the answer plain.

Step 1 of 2

What the top two rungs actually are

The mechanism the table doesn't show
Cloud agent
Spins up a Cursor-managed VM with its own desktop and terminal, checks out the repo, implements the work and opens a PR - running while your laptop is closed. Same models and same harness, so the same quality. The VM is torn down afterwards, so the checkout doesn't persist, and nothing runs on your RAM or disk, so you can have many in flight at once.
Agent SDK
Programmatic access through one runtime-agnostic interface - you pass the runtime (local, Cursor Cloud or self-hosted) and can attach MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. servers, skills and hooks to a run. It's a superset of a skill or a command: it takes the agent anywhere you can run code.
The routing rule: foreground vs background

Foreground (local) is for decisions: interactive, ambiguous, security-sensitive or steered work where you watch every step. Background (cloud) is for execution: clear goals and acceptance criteria.

The decision test: if the output is reviewable through an artifact (image, recording, logs, diffs), push it async; if it's fuzzier, stay foreground. Foreground = exploring an API, a spike, a UI tweak you watch. Background = a refactor, migrating a branch, a database migration.

Anti-pattern: spinning up a background agent with a one-sentence 'refactor the whole codebase' prompt. Background work needs heavy upfront structuring - pull the specific tickets via MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. first.

ZDR caveat - verify before quoting

Zero-data-retention terms and Privacy ModeCursor's setting that guarantees code data is not used for training by Cursor or its model providers, and that an admin can enforce org-wide; data-retention terms are a separate, contractual layer. Press Enter for the full definition. are real enterprise controls. The catch: ZDRZero Data Retention. A contractual guarantee that the model provider won't store your code or train on it. Press Enter for the full definition. does NOT apply when developers use their own API keys.

If a team routes the SDK or agents through personal keys, you've quietly stepped outside the guarantee. Make key management part of the autonomy conversation.