Cowork Scheduled Tasks: Agents That Run With No Device On
Cowork stopped being a desktop-only workspace on July 7, 2026. It now runs on web (claude.ai), iOS and Android, and — the part most write-ups bury — sessions execute on Anthropic's servers, not your machine. That single architectural change is what unlocks the headline feature: scheduled tasks that run with your laptop closed, your phone in your pocket, and no device online at all.
This page is the practical field guide: what actually works on which surface, how to set a scheduled task up correctly the first time, and the small handful of gotchas you only notice after your 6 AM Monday briefing silently fails because it needed a local file.
- Understand why moving Cowork execution off your device changes the product (scheduled runs, cross-device handoff, mobile approval prompts) — and where the tradeoffs land
- Pick the right surface (desktop / web / mobile) for a given job, using the capability matrix
- Set up a scheduled Cowork task correctly on the first try, with the right approval mode, frequency and working folder
- Design a mobile approval flow that doesn't silently stall when you're asleep or on a plane
- Avoid the four gotchas that break scheduled runs in production: local-only tools, Fable-5-on-desktop-only, Dispatch vs Remote, and connector auth drift
The one shift that matters: sessions moved to Anthropic's servers
Before July 2026, a Cowork session was a process inside the Claude Desktop app on your Mac or PC. Close the laptop, session dies. Kill the app, session dies. This is why the pre-July Cowork loop always started with "open desktop, keep it awake."
The July launch introduces remote sessions: the Cowork loop lives in Anthropic's cloud, your session state and files persist against your Claude account, and any surface (web, mobile, desktop) is just a viewer + input device for the same underlying session.
Two direct consequences:
- Scheduled tasks are possible. A recurring task can wake up at 06:00 Monday, do 45 minutes of work, and drop a draft in your account — with zero devices online, because there's no device in the loop.
- Cross-device handoff is free. Start a task at your desk, close the laptop, and see the same session mid-run on your phone. There's nothing to sync — the session was never on your laptop.
The tradeoff is the capability matrix below: anything that needs to touch your local machine (files on disk, your browser, your OS) can only run on desktop, because that's the only surface with a foot in both worlds.
Capability matrix by surface
| Capability | Desktop app | Web (claude.ai) | Mobile (iOS/Android) |
|---|---|---|---|
| Cowork sessions (interactive) | ✓ | ✓ (beta) | ✓ (beta) |
| Scheduled Cowork tasks | ✓ (create + monitor) | ✓ (create + monitor) | ✓ (monitor + approve) |
| Local file access | ✓ | ✗ | ✗ |
| Browser use (local browser automation) | ✓ | ✗ | ✗ |
| Computer use (OS control) | ✓ | ✗ | ✗ |
| Live artifacts (in-session preview) | ✓ | ✗ | ✗ |
| Connected tools (Gmail, Slack, GDrive, Calendar, web) | ✓ | ✓ | ✓ |
| Installed Skills | ✓ | ✓ | ✓ |
| Plugins | ✓ | ✓ | ✓ |
| Mobile push for approval prompts | (via linked phone) | (via linked phone) | ✓ (native) |
| Fable 5 model in Cowork | ✓ (latest Desktop only) | see gotchas | see gotchas |
:::info The rule of thumb If the task only needs connected tools + Skills + web (email triage, "read these three docs and draft me an update", weekly report from Google Sheets), it runs anywhere — schedule it. If it needs your files, your browser, or your machine, it stays desktop-and-live only. :::
Interactive Cowork vs a Scheduled Task vs Dispatch
The three modes get conflated. They're different beasts:
- You start it, you're watching, Claude works over a long horizon on your active goal. Lives on Anthropic's servers now, so you can hand it off between devices. Best for: multi-hour work you want to steer as it happens.
- You define a prompt + a cadence (hourly / daily / weekdays / weekly / manual). Each run spawns a fresh Cowork session with the tools, skills and plugins you specified. Runs remotely, on schedule, whether or not any device is online. Best for: recurring prep work — Monday 06:00 briefings, hourly inbox triage, end-of-day rollups.
- Legacy shape: fire off a background task from an active desktop session that keeps running while you continue chatting. Requires the desktop app to stay awake because the process still lives locally. Best for: local-machine work you want to background while you keep collaborating. Do NOT confuse it with a scheduled remote task — Dispatch dies when the app closes.
The mental shortcut: Dispatch = local background, Scheduled Task = cloud on a cadence, Interactive = you're at the wheel.
Setting up a scheduled task, the right way
A scheduled task needs six things — five settings and one prompt. Get any of the settings wrong and the run will look "successful" but do the wrong thing.
- Short, human. This is what you'll see in the scheduled-tasks list and in mobile notifications. 'Monday client prep' beats 'briefing-v2'.
- Write it as if talking to a colleague who joins fresh every run — because every run IS a fresh session. State the goal, the sources (which connectors, which skills, which folders), the output shape you want, and any invariants (e.g. 'never send email, always leave as draft').
- Pick one of hourly / daily / weekdays / weekly / manual. There is no cron syntax and no sub-hourly options — if you need 'every 5 min', you're on the wrong tool (use the API + your own scheduler).
- The critical setting. Options are auto-approve (Claude proceeds without asking, good for read-only tasks), ask-on-mobile (Claude pauses at a decision point and pings your phone — safe default for anything that writes), or hold-for-review (task always ends with a draft you review later).
- Defaults to the Cowork default for your plan. Pick a smaller/cheaper model for heavy read-summarize loops; pick top-tier only when reasoning quality actually moves the outcome.
- Scopes the session's file writes to a specific project/folder in your Claude account. Skip it and outputs land in a default location — which is fine for one task, a mess when you have twelve.
Scheduled task prompt template: Monday 06:00 client briefing
You are preparing my Monday morning brief for the Acme account. Every Monday at 06:00 local: 1. Read the last 7 days of email in the "Acme" Gmail label. Group by thread; note anything that expects a reply from me. 2. Pull calendar events with "Acme" in the title or attendees from the coming week. Note prep needed for each. 3. Search the web for news about Acme (company name + "acquisition", "layoffs", "product launch") in the last 7 days. Cite sources. 4. Read the Google Doc "Acme - Running Notes" for open questions. Produce a single briefing doc titled "Acme brief — <this-monday-date>" in the "Client briefings" project folder. Structure: - 3-bullet TL;DR at the top - "Needs a reply from me" list (with links) - "This week" list (calendar + prep) - "News since last brief" (with sources) - "Open questions" (from the running-notes doc) DO NOT send any email. Leave a follow-up email as a DRAFT in Gmail titled "Acme weekly check-in" pre-filled with the TL;DR — I'll edit and send.
Two things to notice about this prompt:
- The invariant ("DO NOT send any email... leave as DRAFT") is spelled out. In a scheduled run you cannot ask a follow-up "are you sure?" — the invariants have to be in the prompt.
- The output location + naming is precise. When these run weekly for months, you want them findable, not scattered under default titles.
Mobile approval flow — designing for "I'm asleep"
ask-on-mobile sounds friendly until it fires at 03:00 and the run silently stalls waiting for you. Two rules make this reliable:
- Read-only or draft-only actions → auto-approve. Anything that sends, publishes, deletes, or spends money → ask-on-mobile with a fallback in the prompt: 'if I haven't responded within 30 minutes, save as draft and end the task'. The prompt-level fallback matters because the platform will not time you out for you.
- In the prompt, wrap decision points in a specific ask: 'Before sending anything, request approval with the exact subject line and first 200 chars of the body.' Push notifications truncate — front-load the important bits so you can approve/deny from the lock screen without opening the app.
- Push approvals block the task until you answer. If you never answer, the run sits — it doesn't automatically fall through to a safe default. Bake a 'if no response, do X' timeout into the prompt itself, or use hold-for-review instead.
The four gotchas that break scheduled Cowork runs
- If the prompt asks the task to 'open my downloads folder', 'automate my browser', or 'take a screenshot of my screen', the remote session has no way to do that. The run won't crash — it'll just skip or improvise. Fix: audit your scheduled prompts and swap local-machine actions for connected-tool equivalents (Google Drive instead of local files, web fetch instead of local browser, etc).
- As of the July 2026 launch, running the Fable 5 model inside a Cowork session requires the latest Claude Desktop app. If you schedule a task that pins Fable 5 and rely on it running remotely, verify current support — this constraint is expected to relax but check before pinning.
- Backgrounded Dispatch work from an active desktop session dies when you close the app. If you actually need 'run this weekly, no matter what', don't Dispatch — create a scheduled task.
- OAuth tokens to Gmail, Slack, Google Drive can expire, get revoked, or hit permission changes. A scheduled task under an expired connector produces empty output and no obvious error. Fix: for anything that runs weekly or less often, add a first-step self-check in the prompt ('confirm you can read the Acme Gmail label; if not, produce a one-line error at the top of the doc and stop').
When to reach for the API instead
Scheduled Cowork tasks are a product feature — great for personal / small-team recurring work you want visible in your Claude account, with human approval in the loop. They're the wrong tool when you need:
- Sub-hourly cadence — the smallest scheduled-task interval is hourly.
- Programmatic orchestration — dozens of parallel runs, dynamic scheduling, or tight integration with your infra. Reach for Managed Agents, or roll your own with the Agent SDK + a cron/queue you already run.
- Server-side memory across runs — Cowork tasks pass state via files in your account, which is fine for docs but coarse for structured state. For structured cross-run memory, use Managed Agents Memory Stores.
- Deep multi-agent coordination — for many-agents-collaborating, see Cowork & Agent Teams and Native Multi-Agent APIs.
Quiz
Check yourself
0/4Sources & further reading
- Anthropic — Claude Cowork on web and mobile: hand off work anywhere (launch post, July 7, 2026)
- Anthropic Help Center — Schedule recurring tasks in Claude Cowork (canonical settings + frequency reference)
- Related on AILmanac: Cowork & Agent Teams · Claude Code Background Tasks & Scheduling · Managed Agents · Managed Agents Memory Stores · Connectors