Cursor Guide

Choose the work before choosing the agent

Cursor combines an editor with an agent that can inspect files, edit code and run commands. Its Agent overview describes those capabilities; it does not establish how reliably a particular project will work.

A useful first task is a bounded change with an observable result: repair a failed form submission without losing the user's input, for example. Write the expected behavior and the files that may change before starting. Keep a person responsible for reviewing the result.

Local and cloud work have different boundaries

Cursor's Cloud Agent documentation describes remote virtual machines with repositories, dependencies and network access. The cloud workflow can continue without the developer's laptop staying online. A remote environment still needs the tools and services required to run the project.

The setup documentation explains saved environments and repository configuration. Check which environment is actually selected, which dependency installation ran and whether the relevant tests can execute there. A missing service should produce an explicit acceptance gap, rather than a successful verdict based on a build alone.

Check access and stored data before using client code

The security overview describes cloud agents automatically running terminal commands, and explicitly discusses prompt-injection risk. It also distinguishes workspace, snapshot, conversation and credential retention. Privacy Mode is not a claim that no code or conversation is stored.

The network reference lists unrestricted, default-plus-allowlist and allowlist-only access. It also documents service and source-control exceptions to the allowlist-only mode; the label does not mean every unlisted domain is blocked. Check the active environment's policy, available repositories and required secrets. Start an evaluation with synthetic data and a noncritical repository. This is a proposed evaluation boundary; no client repository was connected for this guide.

Local approval settings should not be assumed to apply to cloud runs. Cursor's Run Modes reference explicitly distinguishes local agents from Cloud Agents.

Budget for the selected model and account

Pricing checked 19 October 2026 lists Free, Plus at $8/month, Premium at $30, Super at $100 and Ultra at $200, before applicable taxes. Check current pricing before purchasing.

The models and pricing documentation says eligible Cursor-model usage, including Cloud Agents, draws from a weekly pool shared with other supported uses. Third-party models on Premium and above use separate prepaid credit from the first request. The Cloud Agent overview still describes API-priced model usage. These pages use different billing descriptions. Before a pilot, check the selected model, remaining pool, extra-credit balance and spending cap in the actual account. No account-specific entitlement or cloud-run cost was verified for this guide.

A small pilot that produces a decision

This is an editorial test plan, not a completed Cursor test. Use a synthetic form with three requirements: accept valid input, reject incomplete input with an explanation and retain the draft after a simulated service failure.

Record the starting commit, allowed changes, selected model, spending cap and exact acceptance commands. After the run, inspect the diff and repeat the normal, invalid and failed-request paths. Check keyboard use and a narrow screen if the change affects the interface. Save failures and skipped checks alongside successful output.

A proposed decision rule is to accept the pilot only when the behavior can be independently reproduced, unrelated files remain intact and the observed cost fits the agreed cap. If an environment cannot run an acceptance check, leave that part unverified. A screenshot, model verdict or passing build cannot answer every requirement.

For a comparison of the work itself, use the same task and acceptance criteria in the coding-agent workflow guide. For help defining a bounded implementation pilot, describe the workflow and the result you need.

Evidence and limits

This is a source-researched guide checked 19 October 2026. No paid Cursor session, performance comparison, product benchmark or hands-on result was collected. There is no score or verified affiliate commission. Availability, billing and configuration for a specific account still need confirmation before a pilot.