Lesson 9.89.8 The same recurring work on today's tools

In everyday life. Two banks both let you set up a recurring transfer. One sends it from a joint account that you share, the other from your own account alone. The transfer is the same. Who it comes from is not.

What you will learn. How each AI vendor's own pages say you can run the same recurring work, and the three differences that change your setup. You practice it in Task 9 of the lab.

Why it matters. The design stays the same on both AI vendors. Who the work runs as, what can start it and how its initiative is limited do not, so the same design can behave differently.

The design in 9.1 to 9.7 is the same on either AI vendor: Anthropic, which makes Claude, or OpenAI, which makes ChatGPT. It is also the same on open-source tools that you run yourself. The products that support it are different.

The comparison. Three differences change your setup.

Identity: a ChatGPT team task runs under a team service account. Claude Tag, a shared Claude in a Slack channel,¹ runs under its own identity there. But a Claude Code routine, an automated job, acts as you.² Claude Duet runs on your Claude account and, with your approval, sends email from the accounts you connect.³ Put team work under a team or worker identity.

Events: ChatGPT's documentation describes tasks started by a new Gmail message. Anthropic documents two kinds of trigger, neither from email: Claude Code routines, from an API call or GitHub, and Claude Tag, from Slack.

Initiative: a dot is OpenAI's always-on agent.⁴ Its own research is read-only, and some of its actions have Custom Rules. But Pause does not stop its scheduled work. Claude Tag is controlled by admins' channel, tool and data settings, and by its "Respond automatically" setting. Claude Tag's pages do not describe a rule for each kind of action. On either AI vendor, write your own initiative rules. Then make them hold by limiting what the worker can access.

Open source: like ChatGPT, OpenClaw, an open-source assistant,⁵ can start work from Gmail. When you run the tool yourself, you set the worker's identity and limits: the accounts and permissions you give it.

What stays the same. Whichever AI vendor or open-source tool runs the work, design as if a run starts fresh. It sees only what it is given, and its trigger can fire twice. Continuity is kept in records the run reads, never in a chat. Initiative is graded on the ladder of 9.6, and it never widens authority. And every piece of work that runs without you has a named person who reads what it did.

For reference.

The two AI vendors' lists use the same seven lines, in the same order. Anthropic's adds one, for its shared worker in Slack.

Anthropic, as verified 8 October 2026.

• Schedules. Scheduled tasks in Claude run at set times, on paid plans.⁶ They run remotely. As of 3 November 2026, two Anthropic pages differ on a task that needs files or apps on your own computer. The scheduling page says it runs only there.⁶ The page on web, desktop and mobile says new scheduled tasks run in the cloud, and a cloud session reaches your files only while the Claude Desktop app is open.⁷ So check where your task says it runs. Chat and Cowork are one Claude, which is being released in stages, starting with the Pro, Max and Ultra plans, as of 31 October 2026.⁸ Claude Code, the coding harness, has routines, in research preview, that add custom schedules.²
• Event triggers. Claude Code routines start from an API call or from GitHub events.² Anthropic's pages do not describe a task in the Claude app that starts when an email arrives.
• Each run. Each scheduled task runs as its own session.⁶ If a routine's trigger is sent again, it starts another session.⁹ A routine's green status means the system running it had no error. It does not mean the task succeeded.²
• Whose name. A Claude Code routine belongs to your own account and acts as you.² In a channel, Claude Tag, a shared Claude in a Slack channel, works under its own identity, with the access an admin sets for that channel. In a direct message, or with a member's own connectors, it acts as that member.¹⁰ Claude Duet runs on your Claude account, and with your approval it sends email from the accounts you connect.³
• Approvals. A scheduled task has an approval mode, from manual approval to "Skip all approvals." Even with "Skip all approvals," permanent file deletion always needs your permission.¹¹ Claude Code routines run without asking for approval, except when they publish some artifacts. So you control them by what they can reach.²
• Standing worker, verified 31 October 2026. Claude Duet is Anthropic's standing worker, in research preview, with a waitlist for Pro, Max and Ultra subscribers in the US. It checks your connected accounts every few hours and sends a short brief each morning. It messages you only when something needs your attention. Before it sends a message, makes a payment, signs a document or deletes anything, it waits for your approval.¹²
• Shared worker in Slack. Claude Tag, in beta on Team and Enterprise, is one shared Claude in a Slack channel.¹ It follows up on its own, and it can reply without being tagged, but only while "Respond automatically" is on.¹³
• Review and stop. Past runs are on the Scheduled page, where a task can be paused.⁶ Claude Tag's Activity page lists every task for Owners, and its spending can be limited.¹

OpenAI, as verified 8 October 2026.

• Schedules. Scheduled tasks run once or on a schedule.¹⁴ How often depends on the kind of task: a task scheduled inside a chat can repeat at intervals measured in minutes.¹⁵ Your plan limits how many tasks can be active.¹⁴ Team tasks, on Business and Enterprise, are recurring work owned by a team.¹⁶
• Event triggers. Tasks in ChatGPT can start from Gmail, Slack or GitHub activity. On Enterprise and Edu, an admin must allow this first.¹⁴
• Each run. A scheduled task that is not inside a chat starts a new chat each time it runs. A task scheduled inside a chat returns to that chat.¹⁵ A team task keeps a list of previous runs.¹⁶
• Whose name. Team tasks run under the team's service account, an account that belongs to the team, not to the creator.¹⁶
• Approvals. Approval rules in connected apps still apply, and an action that needs approval may pause a task.¹⁴
• Standing worker. A dot is an always-on agent with its own cloud computer and memory.⁴ It is being released to Pro and Business Premium, and it is a beta on Enterprise.¹⁷ Its own background research uses read-only tools, which can look but not change anything.⁴ For some actions, its Custom Rules let you choose "take action without asking," "take action when you say so," "ask before taking action" or "hand off to you."¹⁸
• Review and stop. Activity View shows a dot's work.⁴ Pause stops its current main task, but not every delegated task, and not its scheduled runs. Delete does not undo changes already made in connected apps.¹⁸
• A specialist dot, verified 23 October 2026. A company can set one up for dedicated work, with its own identity, credentials and access to the systems it needs. OpenAI's own early testing included invoice processing. It is starting with focused enterprise pilots, in which its engineers help define the dot's responsibilities and how people review and approve its work.⁴

Open source, as verified 23 October 2026.

• Schedules. OpenClaw, an open-source assistant that runs on your own computer,⁵ has a scheduler called Automations. It runs a job once or on a cron schedule, a timetable that repeats.¹⁹ Omnara, a runtime that you can host yourself,²⁰ has cron triggers that run in the time zone you choose.²¹
• Event triggers. OpenClaw's Automations also cover webhooks, messages that another system sends when something happens, and Gmail triggers.¹⁹
• Each run. Each time it fires, an Omnara trigger can start a new agent from a saved profile. Or it can send input to an agent that is already running.²¹
• Standing worker. OpenClaw is a standing assistant.⁵ Its first setup creates a "Proactive check" job, which runs every 30 minutes by default, in the agent's main conversation.²²
• Running it safely. Run OpenClaw on a dedicated machine, virtual machine or container,²³ and give it only the accounts and permissions it needs. Its docs say to avoid opening it directly to the internet.²⁴