Agents in Code

Profiles and running agents#

An agent profile defines a name, purpose, working instructions, and allowed capabilities. A Code subagent is a running assignment created for a task. The profile can be shared with Chat; the assignment, runtime session, and local work remain part of Code.

When signed in, the shared Agents list supplies the desktop's delegation profiles. Enabled profiles carry their purpose, Personality instructions, access level, and selected skills and connectors into Code. Signing out removes that account roster from local delegation. Local Code itself can still be used without an account.

Create an agent from Code#

In the desktop app, open Settings → Agents while working in Code. Enter an Agent name and optional purpose in the New agent form, then choose Create. This adds a reusable profile to the same account list used by Chat, web, and mobile. The Code delegation controls appear on the same settings page.

Profiles can also be created in desktop Chat or through Agents → New agent on web and mobile. Describe the purpose in ordinary language, then review the profile's instructions and access. The shared Agents guide lists each creation path.

Once the profile is available, ask a Code task to delegate a specific assignment to it. Creating the profile does not launch a run; requesting a run does not automatically create a new shared profile.

Model and access differences#

An agent's preferred Chat model is an account model choice. It does not select an installed Code runtime. Code delegation currently inherits the parent task's runtime route; available models and reasoning levels depend on that runtime.

Read-only profiles can inspect and report. Project write profiles can edit only within the permissions granted to the task. A profile's tools cannot exceed the installed and enabled Code capabilities. The parent task's permission ceiling still applies.

Delegate and review#

Use the composer's Mode, permission, delegation control to choose No delegation, Manual approval, or an available automatic mode. In manual mode, review the request in the Agents rail before it starts. Give each assignment a bounded objective and the evidence it should return.

The Code delegation settings include a concurrency limit from one to four and instructions for the lead agent. At most one delegated agent may edit the project at a time. Runtime support can impose further limits. The Agents work surface shows activity and results alongside the main task.

flowchart TD
  accTitle: Shared profiles and local Code delegation
  accDescr: A shared agent profile supplies instructions to a permitted local assignment. The lead reviews its result before a separate Git action.
  A[Shared agent profile] --> C[Assignment within task permissions]
  B[Lead task and runtime] --> C
  C --> D[Local subagent work]
  D --> E[Evidence returned to lead]
  E --> F[Review changes]
  F --> G[Separate commit or push decision]

Memory, mobile, and cloud agents#

An agent's Personality gives standing instructions. Its Chat home is a separate conversation; it is not the Code task transcript. Linked Chat recall, where available, uses the parent's chosen project and the agent's memory policy. See Memory in Code.

Mobile Remote can display the paired desktop's agent activity and supported approval requests. The phone does not launch a local coding runtime. Cursor Cloud Agents is a separate optional plugin: it runs work on Cursor's service using the user's Cursor account, asks for confirmation before dispatch, and does not approve merging the resulting pull request.