Cursor review — cloud agents, subscriptions and builds

What changed

Cursor says cloud agents can now subscribe to events such as pull requests, Slack threads and scheduled tasks, then wake when something happens. Agents can keep a goal active through longer sessions and automatically follow up on pull requests they created, including CI fixes and bot comments.

The release also adds Custom Modes that keep a selected skill pinned, plus subagents running in isolated virtual machines. A related build update creates ready-to-use copies of a development environment so cloud agents can start with dependencies and setup already prepared. Cursor says builds are included with Cloud Agents at no additional cost.

Why it matters

This moves Cursor from an interactive coding assistant toward event-driven software work. The benefit is less repeated setup and fewer manual hand-offs; the risk is that an agent may continue acting after the original prompt. Teams should define approval gates, branch protections, secrets policy and spend limits before enabling always-on workflows.

Checks before relying on it

• Subscriptions are described as cloud-agent functionality, not a universal desktop feature.
• Confirm which plans include cloud agents, builds and isolated subagents.
• Review repository, Slack and PR permissions; test on a non-production repository first.
• Measure failed runs, review time and compute use rather than assuming “always on” means autonomous success.

Primary source: Cursor changelog — August 19, 2026.

August 2026: from prompt to hosted prototype

Cursor now lets Cloud Agents start without a connected GitHub or other source-control provider, preview the result in a browser, and save the work to a private or internal Origin repository. Cursor says publishing through Vercel still requires a Vercel account. The practical change is a lower-friction prototype path, not a replacement for source control.

Use a disposable workspace, set visibility before sharing, and create the Origin repository only after reviewing generated dependencies, data handling, and access. Do not use this shortcut for secrets, customer data, or regulated production code. Compare Cursor vs. Replit when you need a complete builder workflow.

Primary source: Cursor changelog. Verify plan access, repository visibility, Vercel requirements, and connected-app permissions.

What changed in Cursor

Cursor is moving from an editor with an agent panel toward a persistent work coordinator. Projects keep context across larger jobs, delegate work to subagents, and can continue in the cloud when a laptop is closed. Self-hosted machines keep code, build outputs, and secrets on infrastructure chosen by the team. Cloud agents can also react to events such as pull requests, Slack threads, or schedules.

Who benefits and who should wait

These features are relevant to engineering teams maintaining long-running migrations, CI fixes, or multi-repository work. They are less suitable for a solo user who only needs autocomplete, or for organizations that cannot approve remote execution and third-party connectors. Self-hosting reduces some data-movement concerns but does not remove account, agent, or credential governance.

Our assessment and checks

The differentiator is continuity: Cursor can keep a goal and bring work back for review. Before enabling it, define repository visibility, machine enrollment, secret handling, network egress, approval prompts, and cost caps. Start with a non-sensitive repository, require human review for merges and deployments, and inspect the activity log for every delegated task. Cursor’s changelog is evidence of availability, not an independent reliability benchmark.

Primary source and review note

Primary source: Cursor Changelog. Re-check plan eligibility, beta status, machine requirements, and connector permissions before publication.

September 23, 2026 update: Rollouts and Security Review

Cursor’s official changelog introduces two bots for the last mile of software delivery. Rollouts watches a pull request through deployment and reports per-environment health as verified healthy, regression detected, or inconclusive. It reads the diff and affected systems, writes a monitoring plan as a PR comment, then checks deploy events against logs, metrics and traces. If it detects a regression, it can notify the author and, depending on configuration, open a revert PR for review or hand the finding to a cloud agent; it does not merge or roll back on its own.

Security Review reads each pull request in the context of the codebase and posts one comment focused on exploitable bugs. Cursor lists injection across SQL, command and template surfaces, authentication and authorization bypasses, leaked secrets, SSRF, unsafe deserialization, vulnerable dependency changes and other attack paths. Team rules can define codebase-specific constraints. Cursor says both bots are available on Teams and Enterprise plans, with short-term usage credits for roughly 50 changes on Teams and 500 on Enterprise.

Who should use it and what to verify

Engineering teams with source control, a deployment pipeline and telemetry can use Rollouts as a reviewable signal alongside existing tests and change management. Security Review is a useful additional pass for pull requests, not a replacement for threat modeling, code ownership or an independent security assessment. It is not a fit for teams that cannot connect deployment or telemetry systems, or that need guaranteed detection of every vulnerability. Draft PRs are skipped, and the changelog does not claim complete coverage.

Our assessment

The differentiator is the link between an agent-generated change and its production signals. Start with one non-critical repository, keep human approval for merges and reverts, and compare findings with your existing scanner. Record false positives, missed issues, telemetry gaps and time to investigate. We have not independently tested Rollouts, Security Review, integrations or the reported usage credits.

Read the official Cursor changelog, then review our Cursor profile, Cursor vs Replit comparison, and AI builder security guide. These sources support the documented workflow and plan statements; our rollout advice is editorial guidance, not a security guarantee.

October 6, 2026: Cursor Remote Control reaches local agents from iOS

Cursor’s October 6 changelog says users can now see and reply to local agents running on their computer from the Cursor iOS app. Pairing starts in the desktop app, where the user approves the request; after sign-in, the computer and its local agents appear in the iOS app. Cursor says the feature is on by default except for Enterprise organizations, whose administrators can enable it in organization security settings. It does not require Cloud Agents: the work remains on the user’s computer rather than moving to a hosted runtime.

This is most useful to developers who start a local agent task at a workstation and want to check progress or send a follow-up while away. It is a poor substitute for cloud execution when the computer is offline, asleep, or unavailable: Cursor says the host must stay powered on and online, and its setup guidance calls for keeping the lid open and the computer plugged in. The iOS app is a remote-control surface, not evidence that the task continues independently in the cloud.

Decide whether the local-agent workflow fits your team

Before relying on remote control for a deadline-sensitive task, pair a non-production project, test whether the host stays reachable under your power and network settings, and verify what the agent can access in that workspace. Keep repository permissions and secrets limited as you would for a local agent; remote access does not remove the need to review commands or changes. Enterprise teams should ask an administrator whether the organization has enabled the setting. Cursor’s announcement does not describe a separate mobile approval policy or guarantee availability for every account configuration, so confirm the control in the current app before planning around it.

Our assessment: Remote Control improves continuity for a workstation-based agent workflow, but it does not solve unattended execution or host availability. Teams that need jobs to run when a developer’s computer is off should compare this with Cursor cloud agents and use our AI-generated code review checklist before merging changes. We have not independently tested the iOS pairing or background behavior.

Primary source: Cursor changelog, “Remote control for local agents,” October 6, 2026. Recheck platform availability and organization controls before rollout.

October 9, 2026 Agents Window update

Cursor’s 3.24 Agents Window release adds quick controls for Goal, Queued messages, Debug, Canvas, and the local browser preview above the prompt. It also adds one-click expansion and copying for Markdown tables, and lets users download files linked in chat, including files built by a cloud agent. Cursor’s release notes separately describe more resilient cloud-agent reconnection after sleep or network interruptions.

Who benefits: developers who switch between agent instructions, queued work, debugging, previews, and generated files should find these changes easier to discover. They are interface and workflow improvements, not evidence of better code quality or a performance benchmark. Teams should still review generated files and changes before merging or deploying.

Primary source: Cursor release notes (Agents Window 3.24, October 9, 2026).