Agents

• Admin Agents know when a subagent fell back to a different model. When a custom subagent is set to a model your team's admin settings block, that needs Max Mode, or that isn't valid, the parent agent now learns that the subagent ran on the parent's model instead. Subagents that inherit the parent's model also no longer get blocked on teams that disallow the Fast tier when the parent isn't using Fast.
• Admin Computer use stays on when your team blocks the video review model. Fixed an issue where blocking Gemini 3.1 Pro in team model settings turned off computer use for every Cloud Agent on the team, even when the computer use models were allowed. Now only video review is unavailable, and the agent keeps testing with computer use.
• Broken skills are flagged in the run. When a skill's SKILL.md has invalid frontmatter, no description, or can't be read, the Cloud Agent run now shows a warning naming the skill and how to fix it. Previously the agent quietly skipped the skill or was unlikely to pick it.
• Faster follow-ups after turns that only read files. When a turn only reads or searches files, or sends messages, the next turn now starts sooner. The agent no longer waits on a workspace snapshot that wasn't needed because nothing changed.

Pull requests

• Pull requests open ready for review by default. Cloud Agents, including Project workers, now open pull requests ready for review instead of as drafts, unless you or the task ask for a draft. A new Never open PRs as drafts switch in the Pull requests section of your codebase settings keeps every new pull request you and your agents open ready for review, even when a draft is requested.
• Agent comments on private Origin pull requests link back to the run. When a Cloud Agent posts a comment, reply, or review on a private Origin pull request, the comment now ends with an Open in Web button that opens the agent run, so reviewers can jump straight to the agent's work.

Projects

• Project coordinators keep the task tracker accurate. In Projects that use the task tracker, the coordinator now links each pull request a worker reports to the right task, unlinks workers from finished tasks, and closes Queued or parent tasks once other work has delivered them. Workers now report status to their coordinator.
• Parent tasks in a Project's tracker now follow their subtasks. In a Project's task tracker, a parent task with no pull request or agent of its own now takes the status of its most active subtask (In Progress, then For Review, then Queued) and moves to done once every subtask is complete. The status updates on every change to its subtasks, so the List and Board views no longer show a parent stuck behind its children.
• Project workers can check the task tracker. In a Project that tracks tasks, workers and child coordinators can now read the Project's task tracker to see the status of other tasks, without being able to change it. Workers that are still busy with a follow-up message, or that belong to a teammate, also no longer get unlinked from closed tasks when their turn ends.
• Project workers get short topic names you can find later. When a Project coordinator starts workers, it now names each one for what the work is about, like "Webhook retry backoff", instead of a role, action, or PR number. A worker started from only a PR or issue number is renamed once its first report says what the work is about.

Environments

• Manage an environment's secrets from its dashboard page. Each environment's page in the Cloud Agents dashboard now has a Secrets section where you can view that environment's secrets and add new ones, unless its configuration is locked.
• Admin Admins can turn off personal secrets for Cloud Agents. Team admins can turn on Disable personal secrets in the Cloud Agents security settings. Members then can't add or edit personal secrets, and secrets they already added are no longer given to Cloud Agents. Members can still delete their existing personal secrets.
• Admin Agents without a repository run on your team's shared default environment. When a member of a team starts a Cloud Agent without picking a repository or environment, the agent now runs on the team's default no-repo environment instead of a personal hidden one. Nobody can edit, rename, rebuild, delete, or add environment secrets to the team default, but personal and team secrets still apply to it. An unnamed personal environment with no repositories is locked until you give it a name.
• New environments created by a Cloud Agent now use your repo's environment.json. When a Cloud Agent starts on a single repository with no environment yet, it now boots with the repository's committed .cursor/environment.json instead of ignoring it. The new environment keeps tracking that file, so later agents and builds use it too.
• Agents started on a specific environment build use that build's config. Cloud Agents and subagents started on a specific environment build now boot with the environment.json and settings recorded for that build, instead of the repository's current file or the parent agent's newer environment version. This keeps an agent's setup consistent with the build it starts from.
• Verify an environment build with an agent from the Builds section. An environment's Builds section on the Cloud Agents dashboard now has a Verify with Agent link next to Trigger build. It opens a new agent in a new tab, ready to check that the build works. The link appears when you can trigger builds and your team allows manual builds.
• Pick Origin repositories shared with you when setting up an environment. When you create a Cloud Agents environment or edit its repositories, the repository picker now also lists Origin repositories from other teams' namespaces that you've been granted access to, after your own team's repositories. This lets you build a multi-repository environment across namespaces from the dashboard instead of only through the API.
• Admin Tighter, more reliable restricted network access. Cloud Agents with restricted network egress no longer reach every *.cloudfront.net host by default; only the CloudFront hosts needed to pull public container images stay allowed, so add any other CloudFront domain your environment needs to your network allowlist. Repositories hosted on Cursor now clone faster under restricted egress, and changing network settings while an agent runs now applies that agent's current environment allowlist instead of a stale one.
• Default environments no longer trust two compromised root certificates. Cloud Agents' default environments now distrust the emSign ECC Root CA - G3 and AAA Certificate Services roots after private keys for intermediates under them were exposed. The system certificate store is updated, and on the default image the Java trust store is rebuilt from it, so connections signed through these roots fail instead of being trusted.

Self-Hosted Machines

• Open the live workspace of agents on Self-Hosted Machines. Agents running on Self-Hosted Machines can now be opened from the Cursor desktop app with their files, changes, a terminal, and forwarded ports. Turn on the live workspace when you start a Linux machine. If the machine restarts, opening the agent reconnects to it.
• Admin Send self-hosted runs to a team default pool. Team admins can pick a Default pool in the Self-Hosted section of the Cloud Agents dashboard, and self-hosted runs that don't name a pool or machine go there instead of routing by repository alone. If the default pool serves the run's repository, it takes the run. If not, another team pool that serves the repository takes it, and an any-repo default pool takes everything else. Repository access checks still apply.
• Agents on Self-Hosted Machines wait for a dropped machine to come back. When a Self-Hosted Machine drops its connection in the middle of a run, the agent now waits up to 10 minutes for it to come back before reporting the machine as disconnected. If another machine in the pool is free, the agent moves to it instead.

Web

• New agent transcript on the web. Cloud agent conversations at cursor.com/agents, the pull request agent panel, and side chats now use the same transcript as the Agents Window. Your first message in a new agent shows up right away, images included, and in the agent sidebars on codebase pages the mouse wheel and Page Up and Page Down scroll the transcript.
• Fix with agent on pull request pages goes to the right agent. On cursor.com pull request pages, Fix with agent and similar actions now follow up the agent that opened the pull request, or your most recently active agent on it, instead of whichever agent is newest.
• Agent pages show their details sooner on the web. When you open a link to a cloud agent directly, its header, status, and starting prompt appear sooner.
• Share canvases with your whole organization. The share sheet for a canvas can now open it to everyone in your organization, shown as "Anyone at" your organization, or grant access to other teams in your organization.

Integrations

• Linear sessions link to the agent and its PR, and PRs close their Linear issue. Cloud Agent sessions in Linear now show Open in Web, Open in Cursor, and Pull Request links, and the final reply is a plain summary that no longer appears twice. Pull requests from runs started on a Linear issue now end with a Fixes <ISSUE-ID> line, so Linear links the PR and closes the issue when it merges.

MCP

• MCP tools that need an extra permission now ask you to grant it. When an MCP server refuses a Cloud Agent's tool call because your sign-in is missing a permission the server names, the agent shows a sign-in card that names the tool and the permission it needs, instead of failing. If signing in again can't grant it, the agent tells you an admin needs to approve it. Permissions you grant this way are kept the next time you sign in.
• Fixes to Gmail and Google Calendar tools. When Gmail rate-limits a request, the agent now waits briefly as Google asks and retries the call, instead of having later Gmail calls blocked for a cooldown period. When a Google Calendar call fails because the connection is missing a permission, the agent now tells you to reconnect the account instead of showing a raw permission error.
• Google Sheets calls recover when the wrong account is connected. If a Google Sheets tool call fails because the account in use can't open the spreadsheet, the agent now sees your other connected Sheets accounts and can retry with one that has access.

Fixes

• Fixes to Self-Hosted Machines. When a Self-Hosted Machine drops mid-run, Cloud Agents now reconnect to it or move to another free machine in the pool instead of losing the run.
• Fixes to MCP connectors. Common argument mistakes in Slack, Google Sheets, Airtable, and Jira tool calls are repaired automatically, and signing in to a team-owned MCP server no longer conflicts with your own entry for the same URL.
• Fixes to environment builds. The Trigger build dialog now scrolls so its button stays reachable when an environment has many repos.
• Admin Fixes to Cloud Agents dashboard errors. Team preferences on the Cloud Agents page no longer ask paid, Enterprise, or merged teams to upgrade. When a prepaid balance runs out, the error now names it instead of pointing to on-demand usage settings.
• Fixes to agent stores. Project members can open links to a Project's store, and subagents running on their own machine keep access to their Project store after a checkpoint or finished turn. Store traffic from cloud agent machines no longer counts against your own rate limit, and the Agent Stores page shows the server's error message when it fails to load.
• Fixes to how agents handle Origin pull requests. When an agent checks CI on an Origin pull request, checks that are still running but already failing are now flagged, so the agent can start fixing them without waiting. Merge-conflict notices say "at least N" conflicted paths when the list was cut short, and merging a pull request hosted elsewhere now points the agent to that host's tooling. Commits from subagents of a Slack thread agent now carry the same co-author line, and agents no longer commit a stray file created by the co-author git hook.
• Fixes to app publishing. Confirming a publish no longer fails as out of date when the agent committed outside the app's folder or its machine had shut down; the machine now wakes and the confirm retries automatically. New apps are named after their folder or repository instead of a generic name, and a first publish no longer waits on missing secret values. After publishing, the agent names the repository and folder the app came from and links to it, and no longer says secret changes are waiting in the publish dialog.
• Fixes to Pages. Published pages now show updates on the next load after you republish, unpublish, rename, or change their address. Pages no longer fail to compile when code blocks or comments mention import, and images hold their shape with a blurred preview while loading. Creating a page in a team thread where team sharing is restricted now shows your admin's sharing policy instead of a generic error.

Projects

• Project coordinators can start workers wherever your code lives. A Project coordinator can list your environments and repositories, then start a worker in the environment that holds the code, or directly on a repo using its default setup, even when the Project has no repository of its own. Environments and repos outside the Project need your approval from a card that names them, and approved ones appear in the Project's allowed list, where you can revoke them.
• Admin Project agents ask before using a new machine or environment. For every team, a Project's agents now ask a person to approve the first time they use a Self-Hosted Machine, pool, or environment. Team admins can turn on Automatically allow Projects to place their work in the Dashboard under Cloud Agents → Security. Approvals you already granted still apply, and an approval for a repository also covers its GitHub or Origin mirror.

Agents

• Messages sent as an agent finishes are answered in the same run. If you send a message to a Cloud Agent just as it is wrapping up, the agent picks it up as one more step instead of queuing it as a separate follow-up.
• Start Cloud Agents from a short commit SHA. You can start a Cloud Agent on a GitHub repository from an abbreviated commit SHA of 7 or more characters, like the ones git log --oneline prints, instead of getting a "branch or commit not found" error. Branch and tag names with the same text still take precedence.
• Agents can update pull request descriptions they didn't write. When you ask a Cloud Agent to update a GitHub or Cursor-hosted pull request it didn't open, it now adds its content below the existing description instead of refusing, and it can fill in an empty description. Agents in a single-repository environment no longer need the repository named to create or manage pull requests, and completed pull request actions show as finished in the conversation.
• A clearer agents sidebar on pull request pages. The agent list in the floating dock on cursor.com pull request pages opens right above the input and shows how many agents are working or need attention. When the agent you have open is idle but agents under it are still working, its header says "Waiting on" those agents. The sidebar shows a Retry notice instead of loading forever, or tells you when Privacy Mode (Legacy) blocks Cloud Agents.
• Admin Teammate follow-ups on agents without a repository follow your team settings. Teammates can follow up on a colleague's Cloud Agent that runs on a personal environment with no repository, as your team's Team Followups and Allow VM Sharing settings allow. Before, those agents were read-only to teammates. A teammate who opens or follows up on someone else's agent must also have access to every repository the agent was granted.
• Admin Clearer messages when usage limits block a Cloud Agent. When a usage limit stops a Cloud Agent from starting or taking a follow-up, the error tells you what to do next: turn on on-demand usage, raise your own or your team's on-demand limit, or, on a trial, switch to a Cursor model or Auto. If you switch a follow-up to Auto after a trial limit, the agent continues instead of being blocked again.
• Admin Clearer answers when a team model block turns off computer use. If your team admin has disabled a model that computer use depends on, the Cloud Agent tells you that computer use is unavailable, which model is blocked and what it's used for, and that an admin can re-enable it in team model settings.
• Open Settings with a keyboard shortcut. Press Cmd+, (Ctrl+, on Windows and Linux) on the agents pages to open Settings. The collapsed sidebar no longer pops open when you hover over it; click the Open sidebar button instead. Opening a conversation only collapses the sidebar on windows narrower than 1024px.

Environments

• Rename environments and find config-as-code ones on the dashboard. Rename an environment from its detail page with the pencil button next to its name. Environments defined by a committed environment file show a Config as code badge, and a Config as code switch filters the list to them. When a build fails because the GitHub App can't access a repository, the banner names the affected repositories and links to your GitHub installation settings.
• Environment builds no longer silently use the wrong environment.json. When an environment's tracked environment.json is missing, empty, or can't be read, the build now fails with an error that names the file and repository, instead of quietly falling back to .cursor/environment.json or an older saved config. New environments track the repository's environment.json only when you choose that option.
• Fewer duplicate and overwritten environments. Creating an environment from the dashboard or promoting a personal environment to your team now always creates a new environment instead of overwriting one with the same name or repositories. If a team environment with that name already exists, you're asked to rename yours first. Agents started on the same set of repositories reuse one environment.
• Admin Environment build cache is encrypted with your customer-managed key. For Enterprise teams with a customer-managed key, the Cloud Agents environment build cache is encrypted with a key your customer-managed key controls. If you revoke access to your key, or you have no key in the build's region, environments build without a cache.

Self-Hosted Machines

• Admin Teams that require Self-Hosted Machines stay off Cursor-hosted machines. When a team admin turns on Require Private Workers for All Cloud Agents, Cursor no longer builds or snapshots environments for that team or its members, and the Builds tab says builds are turned off. Agents that already ran on Cursor-hosted machines stop accepting follow-ups and ask you to start a new agent on a Self-Hosted Machine.
• Agents on Self-Hosted Machines handle offline machines better. When a pool machine running your agent goes offline, the agent waits for that same machine to come back and continues there, and moves to another machine only after that. If you start an agent or send a follow-up to My Machine while it's disconnected, you get a machine-offline message after about 15 seconds instead of waiting for minutes.

Integrations

• Jira automation rules and GitHub Enterprise @cursor comments launch Cloud Agents correctly. Jira automation rules start a Cloud Agent with the rule's prompt as its task, including rules with no work item attached. @cursor comments on GitHub Enterprise Server repos now work, and run under the team that owns the GitHub app when you belong to it.
• Azure DevOps repos work as org/project/repo shorthands. You can point a Cloud Agent at an Azure DevOps repository with a three-part shorthand like my-org/My Project/portal, for example when switching repos or setting a default repo in Microsoft Teams, instead of pasting the full dev.azure.com URL.
• Admin Cloud Agent commits in OpenTelemetry Export. Enterprise teams using OpenTelemetry Export receive a cursor.cloud_agent.commit log for each commit a Cloud Agent pushes to a GitHub repository. Each log includes the commit SHA, the repository URL, the number of files changed, and line counts when available, and links back to the agent run.

MCP

• Admin Cloud Agents follow your team's plugin and MCP allowlist. When an admin turns on Team list enforcement, Cloud Agents connect only to plugins and MCP servers on the Team list, and the tool restrictions set for each server apply. When enforcement is off, those tool restrictions no longer block other MCP servers.
• Cloud Agents no longer get stuck on sign-in cards for command-based MCP servers. Cloud Agents no longer pause on an MCP sign-in card for stdio servers, which run as a command on the agent's machine and have no Cursor sign-in. Before, the run waited until you clicked Skip or the card timed out. Cloud Agents also no longer ask you to sign in to a server that is still loading.

Fixes

• Fixes to agent runs. Stopping an agent while it continues after a stop hook now works, parallel file writes in one step no longer overwrite each other, and agents no longer get duplicate AGENTS.md or CLAUDE.md instructions. Follow-ups sent during a CI-fix turn run as you, and Bugbot's Fix in Web continues the existing pull request instead of opening a new one.
• Fixes to environment builds and network allowlists. Environment setups longer than an hour no longer fail partway for teams using PrivateLink. A Cloud Agent whose VM is recreated on a follow-up boots the environment's latest successful build, and changing a team environment's egress settings triggers a rebuild. One malformed network allowlist entry is now skipped instead of breaking or widening the whole policy.
• Fixes to repository access. Cloud Agents on GitLab instances that require Git protocol v2 now start, brief GitLab outages are retried, and failed access checks show GitLab's actual response. Tokens or passwords embedded in repository URLs are stripped before they're stored.
• Fixes to Project coordinators and workers. Workers that a coordinator starts with a variant of its own model keep the coordinator's context window instead of jumping to 1M. Coordinators with more than 1,000 workers list their newest workers, and renaming a Project before its first turn now sticks.
• Fixes to Self-Hosted Machines. Runs that land on a machine that is shutting down start on another machine in the pool, and pool machines left behind by a deleted run reset so the next agent can claim them. Agents recover from a lost machine connection by reconnecting on the next step, and Windows machines resolve skill, plugin, and workspace paths correctly.
• Fixes to MCP connections. MCP servers that don't need sign-in no longer have their tools disappear between listings, and servers whose sign-in was revoked or expired show as needing authentication instead of exposing cached tools. A newly added server or tool is no longer reported as missing, and an expired server session reconnects instead of failing the call.
• Fixes to Gmail and Slack tools. When Gmail refuses to open a thread for permission reasons, agents read the thread's first message and note that later messages weren't included. Slack tool calls no longer fail when a channel is given as a #name or a pasted channel link.
• Fixes to code tours. Code tours for pull requests in tall stacks now generate instead of failing, and a tour that fails mid-generation ends with an error instead of leaving the viewer waiting.
• Fixes to recording annotations. Large screen recordings now get chapters, a thumbnail, and trim points instead of being skipped, and annotations are no longer lost when another save happens at the same time.
• Fixes to the web app. The agents navigation drawer no longer opens empty at fractional browser widths, and you see all of your pinned agents, not just the first 1,000.
• Fixes to Slack replies. Agent replies in Slack keep dollar sequences like $1, $&, or $$ exactly as written instead of repeating or dropping parts of the message.

Projects

• Agents can message and read agents you @-mention. When you @-mention another agent to a Project coordinator, the coordinator can read that agent's transcript and send it messages. It can also message its own parent coordinator, and the agent that receives a message can reply directly. Access follows yours: the coordinator reaches only agents you mention and can still open.

Agents

• Subagent progress shows up live in the parent agent. When a Cloud Agent hands work to a subagent, the subagent's tool calls and messages now stream in under the parent's task as they happen, instead of appearing only after the subagent finishes.
• Agent links in summaries now open the right page. When a cloud agent's summary links to another agent, the link in Slack, GitHub, GitLab, and Bitbucket now opens that agent's page on cursor.com instead of breaking. On the web project board, agent links in coordinator files now open the linked agent, including its Changes tab when the link points there.

Environments

• Comments and trailing commas in environment.json work everywhere. Cloud Agents now fully support // and block comments and trailing commas in .cursor/environment.json, wherever the file is read. Repos with a commented environment file no longer get the "set up your environment" prompt, and their snapshot and network settings now apply instead of being silently ignored.

MCP

• Cloud Agents handle MCP rate limits without retry loops. When an MCP server rate-limits a tool, the agent pauses that tool for as long as the server asks, keeps using the server's other tools, and is told when it can retry. Repeated Slack rate limits now cap the wait at about a minute.

Self-Hosted Machines

• Project workers on a named Self-Hosted Machine use that machine's repository. When a Project coordinator starts a worker on one specific Self-Hosted Machine, the worker now works in the repository or folder registered on that machine instead of the Project's repository. Coordinators without a repository of their own can now place workers on a named machine too. Targeting a pool instead of one machine returns an error asking you to name a machine.

Slack

• Admin Choose which team your Slack agents run under. If you belong to more than one paid team, you can now pick the Cursor team that agents you start from Slack run and bill under. Choose it when you link Slack, or change it later with the Team for Slack agents setting in Cloud Agents settings. Follow-ups in a Slack thread keep running under the agent's original team, and a team pinned by your Slack workspace takes precedence.
• Slack previews for agent artifact links. Links to a cloud agent's artifacts, including public artifact shares, now unfurl in Slack with the file name and type, plus the length for videos. Previews only appear when the person who shared the link has access to the artifact, and are skipped in channels shared outside your workspace.

Fixes

• Fixes to agent runs and follow-ups. A message sent after pressing Stop now runs instead of staying queued, and answered questions stay answered after a reload. A run no longer shows a PR from another branch as its PR link.
• Fixes to repository access and tokens. Cloud Agents refresh GitHub tokens during long runs instead of hitting expired tokens mid-run. Repository access errors at start now name the specific repository that is empty, denied, or needs connecting.
• Fixes to the Cloud Agents web app. Keys like Ctrl+K and Ctrl+F in the agent terminal now go to the shell instead of triggering page shortcuts. The Connect GitHub card names the repository an agent needs access to.
• Fixes to MCP tool calls. MCP tool calls recover after a server restart instead of failing, and invalid-argument errors now show the specific reason. Google Calendar time-range searches return only events in that range, and Gmail searches of Spam and of very recent mail no longer come back empty.
• Fixes to Self-Hosted Machines. Agents on any-repo pools no longer pick up untracked rules or instructions left in a checkout by an earlier run. When no worker matches your labels, the error now lists the requested labels.
• Fixes to Slack launches. Changing or resetting your default model and settings from Slack now works for accounts linked under a migrated Slack workspace. When an agent started from Slack fails, the ⏳ reaction is now replaced by ❌ instead of staying next to it.
• Admin Fixes to Remote Control. When Remote Control is turned off, your phone now shows it as disabled instead of a generic unavailable error. Members of enterprise teams that aren't linked to an organization can no longer turn on Remote Control until an organization admin allows it.

Environments

• Environment builds are always on. Environment builds can no longer be turned off. Every environment that isn't a draft now gets automatic builds, and Cloud Agents start from the latest build. Environments that previously had builds turned off now show their active build and recurring build failures, and the Trigger build dialog no longer warns that builds are off.
• Fix Build works on builds whose snapshot expired. Clicking Fix Build on a failed environment build whose snapshot has expired now starts the agent instead of failing. The agent boots the latest successful build and works from the failed build's logs. Launching an agent from a Dockerfile build failure also prefills the composer with a prompt to investigate and fix the build.
• Fewer environment build failures from source control hiccups. Automatic environment rebuilds keep running after a temporary source control access failure and stop only after three such failures in a row. Builds that fail because a self-hosted source control server is down are now reported as a source control outage.
• See which environment changes came from the API. In an environment's version history, saves made through the public API are now labeled "API".
• Wildcard allowlist entries include the parent domain. A network allowlist entry like *.example.com now also lets Cloud Agents reach example.com itself, not just its subdomains.

Agents

• Pinned agents follow you across devices. Pinned agents on the web agents list are now saved to your account, so your pins follow you across browsers and devices. The agents sidebar now groups agents into Today, Last 7 days, and an Older section that starts collapsed.
• Start an agent from a closed pull request. You can now start an agent from a closed pull request whose head branch was deleted. The agent works on a new branch from the base branch, and the panel tells you so.
• Clearer errors when an agent can't start. If an agent fails to start because you lost access to one of its repositories, the error now names that repository. When only your Other Models usage has run out, the error reads "Out of Other Models usage" and suggests switching to a named model or raising on-demand usage. When an agent's machine runs out of memory, the error suggests running memory-heavy steps on fewer files or with lower parallelism.
• Faster agent pages on the web. Opening a Cloud Agent conversation on cursor.com no longer refetches changes that are already loading or up to date, and going back to the agents list is quicker because the list is loaded in the background.

Slack

• Slack model picker uses the web model catalog. The Cloud Agents Slack settings model picker now builds its list from the same model catalog as the web picker and keeps your saved default selectable. The Slack model command now accepts any model you can use, instead of a fixed list.
• Set a channel's Self-Hosted pool from Slack. Any team member can now run @Cursor pool set <name> channel or @Cursor pool unset channel in Slack. Every change to a channel's default repository or Self-Hosted pool is announced in the channel.

Fixes

• Fixes to agent runs and follow-ups. Sending a follow-up after stopping an agent early in its first step now works instead of failing with a VM not found error. Switching between Best-of-N tabs now loads the selected candidate's conversation. Agents started on a saved environment that changed after its last build no longer fail at startup with a build version mismatch.
• Fixes to multi-repo environments. Cloud Agents now start normally even when one of your multi-repo environments has colliding clone paths; before, that one environment made every start fail. Saving an environment whose generated name lists many repositories no longer fails.
• Fixes to the Cloud Agents dashboard. If the Cloud Agents settings page can't load privacy settings, it now shows an error with a Try again button. Failed loads on the runs pages also offer Try again, and the runs summary reads "1 run" instead of "1 runs".

Environments

• Activate a draft environment build while it's still running. You can now click Activate on a draft build that is still in progress, and it becomes the active build for your environment as soon as it finishes. Activating a draft whose snapshot has already been deleted is still blocked, and no longer changes your saved environment settings.
• Default-on environment builds work like turned-on builds. Environments with no build setting of their own, where builds are on by default, now show their Active Build and the failing-builds warning, and are rebuilt when a build is triggered, so agents in them start from a fresher build.

Agents

• Long-running mode removed from the web. The Long-running option in the web model picker, its time-limit pill, and the team Long-running agents setting are gone. Cloud Agents started from the web always run in standard mode.
• Coordinators can switch a worker's model. A Project coordinator can now move a worker to a different model when it sends the worker a message. The new model applies from the worker's next turn.

Self-Hosted Machines

• Shorter self-hosted options in Slack, GitHub, and Linear. Add sh=t to a Slack, GitHub, or Linear request to run the agent on your self-hosted pool. self_hosted also accepts t/f and 1/0, and all three surfaces now take the same options. Options written inside Slack code blocks no longer start or configure an agent.
• Take control of desktops on Self-Hosted Machines. Taking control of a Cloud Agent's desktop now reaches the machine you're viewing, including agents on Self-Hosted Machines. Older machines fall back to the previous takeover.
• Agents on your own machines handle git and setup more carefully. Cloud Agents on Self-Hosted Machines now check the machine's git name and email before committing, and set a repository-only fallback only if none is set. They ask for credentials on the machine instead of assuming Dashboard secrets or a signed-in GitHub CLI. On My Machine, agents prefer project-local setup over system-wide installs and global config changes.
• Slack uses a channel's default Self-Hosted pool. @Cursor launches in Slack now use the channel's default Self-Hosted pool when one is set, then the team default. If a launch is rejected, the reply names the channel default and how to change it with @Cursor pool set <name> channel.

Models

• Grok 4.7 on the Start plan. Start plan users can now pick Grok 4.7 for Cloud Agents, fixed at medium effort and standard speed.

Reliability

• Clearer errors when image generation fails. Image generation failures in a Cloud Agent now show a readable explanation instead of a raw status code.
• Fixed moving some local chats to the cloud. Moving a local chat to a Cloud Agent without typing a new message no longer fails with a "Missing required fields" error when the chat's saved state has no turns. The Cloud Agent now starts idle, keeps the chat's history, and runs your next follow-up.

Fixes

• Fixes to environments and agent starts. Starting a cloud agent no longer overwrites the environment's saved repositories, and boots the default branch when no branch is given. Failed environment builds now report the specific cause (clone, source-control access, or pod start), and agents started with agent-scoped secrets now boot from the environment's build.
• Fixes to Self-Hosted Machines. Agents and subagents started on a self-hosted pool that accepts any repository now get picked up instead of waiting forever, and idle or deleted-run machines free up right away instead of showing as busy. Agents on Origin repositories that were detached from GitHub now receive their configured secrets.
• Fixes to MCP connectors in Cloud Agents. Connector tools that Cloud Agents list now accept the calls made to them, including for connector accounts with custom labels. Plugin-backed MCP servers that an admin granted now serve their tools, even when the dashboard label differs from the plugin name or the agent runs as a deployment service account. The environment info tool now shows every documented environment.json field, and for subagents it reports the network egress policy that is actually enforced.
• Fixes to agent runs and follow-ups. A follow-up sent while an agent's machine is being recreated is now saved to the queue immediately, so it no longer looks lost until the new machine is ready. An agent's link to its Origin pull request no longer disappears after a temporary lookup failure. Agents started or followed up with a specific model variant are now checked against your team's model policy using that variant, so they're no longer wrongly blocked. An agent left stuck behind a queued run that never started accepts follow-ups again after an hour instead of rejecting them.
• Fixes to artifacts in agent pull requests. Artifact links that an agent puts inside tables or other HTML in a pull request description now render as working links instead of raw text. Pull requests no longer get public artifact links for files that couldn't be confirmed to exist, and agents with many artifacts no longer re-announce the same artifacts as new on later turns.
• Fixes to Slack message formatting. Agent replies in Slack keep links whole, including URLs with parentheses or & and link text with brackets. Bold italic text, ~~~ code fences, task-list checkboxes, and images (now shown as links) render correctly, and text like __init__ or 2*3*4 is no longer mis-styled.
• Fixes to the Cloud Agents page and dashboard. On cursor.com/agents, the first page of agents is now fetched on the server while the page renders, so it shows up faster on a full page load. In machine settings, the remote control toggle now explains why it is off and who can turn it on. Runs summary stat cards also reflow to fit the available width.

Agents

• Follow-ups to agents archived elsewhere now go through. If you archive an agent on another client, such as the desktop app, and then send a follow-up from a tab you already had open on the web, Cloud Agents now unarchives the agent and delivers your message instead of failing. This works from the agent chat, resubmitted messages, side chats, and the review agents sidebar on pull request pages. Deleted agents stay closed.
• Plan and limit dialogs appear where you start the agent. When a Cloud Agent can't start because of your plan, usage pricing, model choice, or a protected scope, the dialog explaining why now opens on the page you started from, for pages within the agents area of cursor.com, instead of later on the agents page.
• Draft pull requests stay in draft. Fixed an issue where Cloud Agents marked their draft pull requests ready for review on their own, even when your repository rules said to keep them as drafts. Agents now follow your instructions and repository guidance, and still mark a PR ready when you ask.
• Fixed agents reading the default branch when started on another branch or commit. When you start a Cloud Agent on a specific branch or commit, it now waits until that ref is checked out before running any tools, so it no longer reads files from or describes the default branch first.
• Cloud Agents keep their environment and files when it is moved to another machine. Previously, if an agent's environment was moved while you sent a follow-up, the agent could report that the environment was unreachable and continue in a fresh one, losing your files. Now the agent shows "Reconnecting to your environment. This can take a few minutes.", retries on its own, and continues in the same environment with every file in place. If reconnecting takes too long, it asks you to send your message again instead of starting over.
• Pin up to 75 agents. The agents list on cursor.com now lets you pin up to 75 agents, up from 50.
• Blocked models are now rejected as soon as you send a follow-up. If you switch a Cloud Agent follow-up to a model your team admin has blocked, you now get a Model Blocked error right away. Before, the follow-up was queued and then failed mid-run or quietly ran on a different model.
• Fixes to page layout on the Agents web app. Opening a conversation, repo, or pull request no longer makes the page jump when the sidebar collapses, and the sidebar loads already collapsed if you keep it that way, or when you open a conversation in a narrow window. The agents list no longer shifts sideways when its scrollbar appears, or moves down when promotions, the Cursor Learn banner, or the connect source control prompt show up after the page loads.
• Fixes to Cloud Agent runs and follow-ups. Queued follow-ups, repo skills, and slash commands now load in side chats, and repo command pills from the composer pass their contents to the agent. PR creation failures from GitHub rate limits now say how long to wait.
• Fewer failed starts and follow-ups. When a Cloud Agent can't get a worker to start its session, it now retries on its own; if it still can't, it says "Cloud Agent could not get a worker to start your session; please retry in a few minutes." Follow-up turns no longer pick up another account's personal rules or skills from the workspace. The slash command menu of a failed or archived agent stops waiting for commands that will never load, and folders deleted from an agent's Context no longer reappear with leftover files.
• Faster starts on a named branch. Cloud Agents started on an explicitly named branch, whether the default branch or another one, can now use the faster startup path instead of falling back to the older start.

Projects

• Team admins' Projects setting now applies to Cloud Agents. When a team admin turns off Projects, the team's Project agents no longer show up in Cloud Agent lists, and starting a new Project fails with "Projects are disabled for your team." Personal accounts, and teams that haven't changed the setting, keep Projects on.
• Fixes to Projects coordinator and worker agents. Coordinators can now start workers after a Project's repository is published under a new name, and PR label edits, environment drafts, and shared links use the published name too. Force-submitting a message to a worker no longer leaves the coordinator's call to it stuck as running. When a worker can't be created, the coordinator gets the real reason, such as hitting the nesting limit, instead of retrying a generic error.
• Cloud workers share the Project's Context. Cloud workers that a Project coordinator creates now use the coordinator's Context as their own, so files they save there are visible across the Project. Moving a worker out of the Project gives it its own Context again.

Environments

• Fixes to environment builds. Builds stuck in progress no longer block scheduled builds, and stuck builds are marked failed after 9 hours instead of 24. The environments list also loads for large teams instead of timing out. Team environment builds that failed because the team's build account couldn't access a repository now rebuild with the access of the person who started the agent.

Self-Hosted Machines

• Self-hosted pools now respect repository access. Starting a cloud agent on a team's Self-Hosted Machines pool now fails with an access error if you can't read the repository. Pool listings only show the repositories you have access to, and API keys limited to certain repositories can only use pools for those repositories.
• Self-hosted agents without a repository can set one up. Agents on Self-Hosted Machines that start without a repository now follow your worker's rules and credentials to find or clone the repository they need, instead of saying they have no repository access. If no rule explains how to get the repository, the agent says it isn't configured rather than guessing a clone URL.
• Fixes to self-hosted and private workers. Agents on pools whose machine is still booting wait for it instead of failing, dead worker connections reconnect or reassign instead of failing with a closed-transport error, and agents on Windows workers no longer break the clients that display them. Subagents sharing a parent's worker no longer orphan the parent's workspace.

Integrations

• More reliable self-hosted routing from Slack, GitHub, and Linear. Mentioning the word self_hosted in a request no longer sends the agent to a Self-Hosted Machine by accident. Opt in explicitly with self_hosted=true, which now also works inside inline code and with trailing punctuation. In GitHub and Linear, options inside fenced code blocks are ignored. In Slack, pool=<name> for a self-hosted pool with no repository now starts the agent on that pool instead of failing to match the team's default repository.
• Jira launches use the right repo host. Cloud Agents launched from Jira with a default repo set as owner/repo now run against that repo on your connected GitLab or Azure DevOps host, instead of assuming github.com.

MCP

• Cloud agents with many MCP tools no longer fail on model tool limits. When your MCP servers expose more tools than a model accepts, cloud agents now cap the MCP tools they send, keeping browser and custom tools first, instead of failing the request.
• Clear error for empty Cursor-hosted repositories. Starting a cloud agent or environment build on an empty repository hosted on Cursor now says the repository is empty and asks you to add an initial commit, instead of showing a generic failure.

Dashboard

• Fixes to Cloud Agents settings and machines pages. The My machines list no longer shows the previous team's machines after you switch teams, and its page is kept in the URL so links and back navigation land on the right page. Closing the New environment dialog in any way, including browser back, now clears the draft instead of leaving stale repositories, name, and scope. Settings sections such as Preferences and Security have shareable anchor links, and counts show thousands separators.

Agents

• The agent steps back when you take control of its desktop. When you take control of a Cloud Agent's desktop, the agent stops using the computer instead of competing with you, and the desktop is released as soon as a computer-use subagent finishes.
• Faster starts when you name the default branch. Cloud Agents started with the repository's default branch named explicitly, as the desktop app does, now reuse the prebuilt environment, the same as starts that name no branch, so they reach the first response sooner.

Projects

• Project coordinators and workers show the right state. A Project coordinator is now marked unread only after a turn in which it actually sent you a message, and the messages it sends you now appear in the agent's conversation history. Workers and subagents that finished stay marked finished when a new message or a resubmitted message interrupts the turn that reports their completion.

Pull requests

• Pull requests from mirrored repositories open in the right place. When a repository is mirrored between GitHub and Origin, Cloud Agents now open the pull request on whichever side is the source of truth and tell you which one they used. If the agent can't create the pull request, it reports the reason instead of working around it with another tool. Pull request links for mirrored repositories now point to the GitHub, GitHub Enterprise, or Origin host where the pull request actually lives, and agents no longer target a commit SHA as the base branch.
• Merge approvals show the exact commit an agent will land. When a Cloud Agent asks for approval to merge a pull request on an Origin repository, the request now shows the commit subject and message the agent chose, in both the mobile approval sheet and Slack. You approve the commit that actually lands in history, not just the pull request title.

Subscriptions

• CI subscriptions on your default branch no longer wake an agent on every merge. When a cloud agent subscribes to CI on a repository's default branch, such as main, it now wakes once for the next finished CI result, and then the subscription closes. To have an agent keep watching a pull request's CI, have it subscribe to the PR's branch.

Environments

• Fixes to environment setup. The environments list loads faster for teams with many environments. The numbered steps in the environment setup intro dialog now line up evenly.

Self-Hosted Machines

• Fixes for Cloud Agents on self-hosted workers. Trying to attach to the workspace of an agent on a self-hosted worker now shows "Workspace runs on a self-hosted worker" right away, instead of waiting on a workspace that never becomes ready.
• Fixes to self-hosted worker claims. Self-hosted workers that reconnect no longer stay stuck on a stale claim, so they can be picked up for new agents again. Claim rejections no longer show up as "No self-hosted workers available", and claimed workers that never received a message are now released after the idle timeout.

Slack

• Set a team default Self-Hosted pool for Slack. Team admins can run @Cursor pool set <name> so @Cursor mentions run on that Self-Hosted pool without adding pool= each time. If the pool serves any repository, mentions that name no repository can still launch an agent there. @Cursor pool and @Cursor settings show the current default, and @Cursor pool unset clears it.
• Fixed Slack follow-ups that reached an agent without the thread. When you mention Cursor in a Slack thread before its Cloud Agent has posted a reply, the agent now receives the whole thread, including forwarded messages and files, instead of only your latest message. Follow-ups sent from the follow-up dialog get the same full context.
• Mentioning @Cursor in another team's agent thread now starts a new agent. When you mention @Cursor in a Slack thread where the agent belongs to a different Cursor team, Cursor now starts a new Cloud Agent for you. Before, it refused and suggested Team Followups, which can't share agents across teams. If a follow-up is still denied, the message now says the agent belongs to another team.
• Fixes to launching agents from Slack. When a Slack launch fails, the reply now names the repository and the real cause, such as no access, repo not found, a disconnected source control provider, or GitHub being unreachable, and links to GitHub Status and a Try Again button for that last case. Mentioning Cursor several times in one thread no longer starts duplicate agents when an earlier launch is slow, and your saved default repo and branch now still apply after your Slack workspace is migrated.

Jira

• More reliable Cloud Agents from Jira. Cloud Agents started from Jira keep working with the latest Jira agent protocol, and show as working instead of paused while they run. Each new request from Jira now starts its own conversation, follow-up messages on an active task no longer fail, and stopping an agent while it starts now pauses it once it launches.

Fixes

• Fixes to Cloud Agent reliability. Stopping an agent no longer shows a spurious provider error, follow-ups no longer fail to restore the branch when untracked files are in the way, and unnamed terminals in environment.json now start. Plugin slash commands in follow-ups now wait briefly for a plugin that isn't ready yet instead of skipping it.

Web app

• Fixes to the agents web app. Clicking an agent in the agents sidebar on cursor.com no longer bounces back to the previous agent, and back and forward restore the tab you had open. A new agent started from the web now shows its logs and conversation once it's created, instead of missing them.

Agent runs

• Fewer pointless wake-ups from subscriptions. Cloud agents subscribed to GitHub CI or pull request events no longer wake for repeated passing CI results, their own pushes and PR edits, or label, title, and assignee changes. A failure, or the first green after one, still wakes the agent. When an event needs no action, the agent now ends its turn quietly instead of posting a filler reply in Slack or on the PR.
• Plugin slash commands work in follow-ups. You can now use a plugin slash command in a follow-up to a Cloud Agent, even if that plugin wasn't selected when the agent started: the agent installs the plugin before it runs the turn. If the plugin can't be installed, the agent says so and asks you to remove the command and send it again, or start a new agent.
• Coordinator messages reach busy workers sooner. In Projects, a message a coordinator sends to a worker that is in the middle of a turn is now delivered into that turn by default, instead of waiting for the turn to end. A worker that isn't running gets it as its next message.

Environments

• Clearer environment setup agents. The first time you open an environment setup agent on the web, a short dialog with a video explains what the agent will do: install and verify your app, ask for secrets or network access, and prompt you to review and save the configuration. Setup agents are now named after the repository, like "Set up my-repo environment", and transcripts no longer show an environment line when the linked environment has no configuration.
• Pin cloud roles to a single environment with OIDC tokens. The identity token endpoint inside the agent's environment now accepts sub_claim: "environment_id", which puts the agent's environment ID in the token's sub claim, so verifiers like AWS STS that only match sub can trust one environment. If the agent has no environment, the mint fails instead of falling back to the default subject.
• Fixed Origin CLI commands failing in Cloud Agents with restricted network access. Cloud Agents with restricted network egress working on repositories hosted on Cursor now automatically allow the Origin API hosts alongside the Git host, so origin CLI and API commands connect instead of failing. Other repositories and your custom allowlists are unchanged.

Self-Hosted Machines

• Fixed pool labels in Linear so runs go to your Self-Hosted Machines pool. A Linear issue or project label whose parent label is pool now sends the Cloud Agent run to the pool with that name, as the docs describe. This also works when you assign or delegate an issue to Cursor, which has no comment to read. Before, those runs quietly used Cursor-hosted machines. Writing pool=<name> in the @Cursor comment still overrides the label.
• Agents can find your Self-Hosted Machines. When you ask a Cloud Agent to run work on a self-hosted machine, it can now list the connected machines you have access to and pick one. Before, the agent could fail with a tool-not-found error that made it look like your machine was gone.

Integrations

• Better Slack previews for Cursor-hosted pull request links. Pull request links from Cursor-hosted repositories now unfurl in Slack with a formatted excerpt of the description, kept short so previews stay compact on Slack mobile. Links to a pull request's tabs, like its changes or commits, now unfurl too.

Notifications

• Fewer noisy push notifications from Cloud Agents. Cloud Agents no longer push a turn-finished notification to your phone or browser. You still get push notifications when an agent asks a question or needs your approval.

Performance

• Faster environment loading. MCP server configs now load in parallel for users with many servers. The environments list also loads faster and no longer times out for large teams with many repositories.

Fixes

• Fixes to environment and repository setup. Saving an environment with a deleted environment's name now creates a new one, and starting an agent with an unknown environment name reuses the repo's existing environment. Multi-repo agents no longer fail on GitLab repos missing from the repo list. Repository checkout now retries on temporary proxy errors, and teams using Customer PrivateLink no longer see clone failures during environment builds or lose git access mid-run.
• Fixes to self-hosted workers. Cloud agents on self-hosted workers no longer fail to start or resume after an earlier failed attempt when idle workers are available, and they no longer fail with a screen recording error. Plugin slash commands and skills now show up when the plugins were already installed on the worker. Coordinator and worker notifications and PR actions also work again on workers with a repository set.
• Fixes to MCP authentication and agents that manage other agents. Cloud agents no longer pause to ask you to authenticate an MCP server that's already connected, and stopping an agent during MCP authentication now closes the Authenticate card instead of leaving it pending. When an agent starts or messages other agents, the start card now completes after the new agent's first turn, and a queued message that gets removed ends as canceled instead of waiting forever. When a subagent can't start because your included usage for its model has run out, the error now names the model. On the iOS app 1.9.0 and later, Cloud Agents that need MCP sign-in now show the sign-in card instead of a fallback message.
• Fixes to cloud agent runs. Cloud agents started from a registry model now run with the effort and parameters you pick in the model picker, not the model's defaults. Cloud agent runs and integration launches now use your account's current privacy mode instead of an outdated setting. Messages you send to a running agent now arrive in the order you sent them, even when an earlier message missed the agent's current step and waits for its next turn.

Web app

• Cleaner Cloud Agents layout on your phone. On mobile, the agents pages on cursor.com use slightly larger text, a centered top navigation without the Dashboard tab, and a logo that takes you back to your agents. Search and the source filter now sit inline with the agents list, the composer has a taller input, and the menu button shows your profile photo.