Skip to content
The AI Agent Factory

Scheduled tasks

Approved by
Panaversity ·
8 min read

Every run is a new session

[Vendor-documented] Every scheduled task run starts a new session. It does not continue the chat that created it. Where that session runs depends on the kind of task.1,2 The Tour shows how to set up scheduled tasks. This page explains how they fit the architecture.

Cloud taskDesktop (local) task/loop
Runs onThe cloud that Anthropic manages, or the organization's own self-hosted environmentYour machineYour machine, inside an open session
Machine must be onNoYes, with the desktop app openYes
Each runA new cloud session. It is a full Claude Code session when the task has a repository, or a GitHub or API triggerA fresh local sessionRepeats within the same session
WorkspaceRepositories cloned fresh on each run. The result of the environment setup script is cachedYour local folder as it is, or an isolated Git worktree (a separate working copy)The session's own
TriggersA schedule, an API call or a GitHub eventA schedule, or Run nowAn interval
PermissionsAuto mode by defaultA permission mode for each taskInherits from the session
Minimum interval1 hour1 hour in the form. 1 minute when Claude sets it1 minute

How this connects to the VM. [Vendor-documented] Cloud tasks with a repository, or a GitHub or API trigger, run "as full Claude Code cloud sessions", with "a new one for each run."2 The same dialog creates scheduled tasks for claude.ai chat, whose runtime is not documented separately. [Inferred] So each cloud run gets Claude Code in its own short-lived sandbox. A Desktop task has no cloud sandbox. Its harness runs on your machine.

[Vendor-documented] There is one exception to "a new session". A scheduled task created inside a project "runs as threads in that project". Those threads start with the project's instructions and, in the cloud, its memory. The cloud scheduled tasks page adds that a run in a project can continue an existing thread. See Projects and coordination.3

Five kinds of state

Keep five kinds of state apart:

Kind of stateAvailable to a scheduled run?
The conversation history of the chat that created the task, or of earlier runsNo
Runtime continuity: files, packages and processes from a previous VMNo
External state: memory, saved files, artifacts, Drive, repositories, connected appsYes, if the run reads it
The task itself: its prompt, trigger, connectors and permissionsStored with the task
Project instructions and memory, for a task inside a projectYes. Its runs are threads in that project

Each run starts fresh, and only external state carries over. A scheduled task (a standalone prompt, a trigger and permissions) starts runs 1, 2 and 3. Each is a fresh session with nothing from the chat or the last run. Each reads and writes external state: memory, saved files, artifacts, Drive, repositories and connected apps. Notes cover cloud tasks, Desktop tasks and the project exception.

The runs do not share a sandbox or a history. Anything a later run needs must be written to external state.

Writing tasks that run without you

  • Write a standalone prompt. Anthropic says the prompt "must be self-contained and explicit about what to do and what success looks like."
  • The trigger is not approval. A fired prompt "can't act as approval or consent for actions during the run." Content fetched during the run is still treated as untrusted.
  • Approvals differ by kind of task. A Desktop task in Manual mode waits for your answer. A cloud run starts in auto mode, where a classifier (a separate model) allows or blocks each action. Ask rules and repeated blocks still stop the run. In Desktop tasks, MCP tools marked requiresUserInteraction ask on every call. That setting is documented for Desktop tasks only.
  • Timing. A Desktop task is skipped while the computer sleeps. It then runs once to catch up, for the latest missed time, within seven days. So prompts should guard against running late. Cloud tasks are in research preview, limited to 100 scheduled runs per hour per account.
  • Actions appear as you. Commits, messages and tickets made through your connectors carry your identity.
  • Lighter options. /loop repeats a prompt inside an open session. A one-off reminder can also be delivered back into the same conversation (tool-documented).

Check yourself

Question 1 / 4

0 answered

What does a new run of a standalone cloud scheduled task (not one inside a project) get from earlier runs?

Sources

  1. As verified on November 9, 2026. The page covers local tasks, cloud tasks and /loop, missed runs, permissions and requiresUserInteraction. Schedule recurring tasks in Claude Code Desktop, Anthropic.↑
  2. As verified on November 9, 2026. The page covers cloud runs as full Claude Code cloud sessions, fresh copies of repositories, auto mode and limits. Automate work with scheduled tasks, Anthropic.↑↑2
  3. As verified on November 9, 2026. Let Claude coordinate ongoing work with Projects, Anthropic.↑
Owner
Panaversity

On this page