Previously, users have needed to BYO devices to host their AI agents. Whether that's running Claude Code et al locally on your machine, running OpenClaw/Hermes on a local server, or setting up a VPS in the cloud, there weren't many out-of-the-box solutions that gave people an always-on, always-available agent host. Recently, things have changed. With the release of Claude Tag, Grok Bot, Meta Muse, and ChatGPT Dots, more and more providers are now offering subscription-based agents you don't need to host yourself. This opens up a lot of new opportunities for AI automation, but where (and how) do cloud-based agents fit into the mix? When is the best time to reach for them over your existing local setup?
Most AI tools you currently use need you sitting at your machine. These are still a first-class choice for your own interactive work, where you'd be sitting at your computer anyway. Cloud agents don't need you at the helm, and if you've been swimming in AI automation for any length of time, you've probably run into scenarios where being tethered to your own device is a hindrance, not a help.
3 questions can now help you decide between local vs cloud agents:
| Attended (you drive) | Unattended (it starts itself) | |
|---|---|---|
| Local | Claude Code, Copilot in your IDE, Codex | Desktop scheduled tasks, /loop dies when the laptop sleeps e.g. OpenClaw, Hermes |
| Cloud | ChatGPT, Claude chat, Claude Code on the web | Cloud agents e.g. Routines, ChatGPT Tasks, Claude Tag, Dots, Grok Bot, Foundry |
Chat and code are for work you want to steer. Cloud agents are for work you want to stop thinking about.
Use cloud agents when the task is:
Some examples that fit well with cloud agents are:
Every Monday at 9am you paste the sprint board into ChatGPT, ask for a summary, and copy it into the team channel.
❌ Figure: Bad example - You are the scheduler, and you forget
A Claude Routine runs Mondays at 8am, reads the Sprint board through a connector, and posts the summary to the channel before standup.
✅ Figure: Good example - The agent is the scheduler, and it doesn't forget
Stick with chat or a local agent when you are still working out what "done" looks like. If you'd change the prompt every run, it isn't ready to automate.
A cloud agent runs on the provider's machine, not yours. It can't see your working folder, your uncommitted changes, your local database, or anything that is only reachable from inside your network. It only sees what you hand it: a repository it can clone, a connector to a cloud service, or a file you upload.
Make this a decision point before you automate anything:
| Where the task's data lives | Use |
|---|---|
| Already in the cloud - GitHub, SharePoint, Slack, a SaaS app with a connector | Cloud agent |
| On your machine, but easy to move - a folder you could push to a repo or sync to cloud storage | Move it first, then use a cloud agent |
| Only on your machine or inside your network - uncommitted work, local databases, internal-only systems, licensed desktop apps | Local agent |
You schedule a cloud agent to summarise the meeting notes in your Documents folder every Friday. It runs on time and reports that it found nothing.
❌ Figure: Bad example - The agent can't reach files that only exist on your laptop
You move the meeting notes to a SharePoint library and give the agent a connector scoped to that library. It reads the same notes every Friday, whether your laptop is on or not.
✅ Figure: Good example - The data lives where the agent can reach it
If the data can't leave your machine, that settles it. Keep the task on a local agent and accept that it only runs while your machine is on.
The first 5 are products you configure in minutes. Foundry is a platform you build on. If you're choosing between Microsoft agent options, see Do you choose the right platform for your AI agent?
| Platform | Starts on | Best for | Watch out for |
|---|---|---|---|
| Claude Routines (Recommended for dev work) | Schedule (hourly min), API call, GitHub event | Code and repo chores - PR review, docs drift, alert triage | Research preview. Runs as you, with no approval prompts |
| ChatGPT Tasks | Schedule (hourly min on paid plans) | Personal reminders and research digests | 3-15 active tasks per plan. Idle tasks can auto-pause |
| Dots (OpenAI) | Chat in ChatGPT, Slack, Teams, or a phone call. Checks in on its own when idle | An always-on assistant that watches your inbox and projects across 4,000+ apps | Rolling out on Pro, Business Premium, and Enterprise. Not yet for Pro users in the UK or EU |
| Grok Bot (xAI) | Chat from desktop or mobile. Learns a job by watching you do it once | Bots that sign into your apps, even ones with no API | Bundled with SuperGrok and Cursor plans, with separate usage. Bots share one cloud computer and your logins |
| Claude Tag | On-demand using Slack | Ad-hoc tasks that don't require direction | Unaddressed feedback or questions that stall progress |
| Azure AI Foundry | Cron (5 min min), timer, GitHub issue, Logic Apps | Production agents that need Azure networking and governance | Pro-code. You build and host the agent |
An unattended agent acts with your credentials and nobody watching. Claude Routines commit as your GitHub user and post to Slack as you. Scope every connector and repo to what the job needs, and read the run output for the first week before trusting it.
A local agent asks before it does something risky, and you are there to say no. A cloud agent has no one to ask, so the permissions you give it up front are the only control you have.
For the wider picture, see Do you manage security risks when adopting AI solutions?