Lesson 3.73.7 How today's tools support the rhythm

In everyday life. You can ask your bank to send money for you on a schedule, such as your rent every month. Banks keep these scheduled transfers safe in different ways.

• The first bank, like Claude. When you set up a scheduled transfer, it asks you to pick one setting for it: ask me before every transfer, check each transfer for me, or don't ask at all. One choice covers everything that scheduled transfer does, as Claude's approval mode covers every action in a scheduled task.
• The second bank, like ChatGPT. It asks you to approve only certain kinds of transfer, such as payments to someone you have never paid before. Everything else goes through. As on ChatGPT, only some kinds of action need you.
• The safest account. Either way, the safest account is one that no scheduled transfer can reach at all. An approval only makes a person decide. An account no transfer can reach makes a wrong transfer impossible.

Both AI vendors let an everyday user hand over recurring work that runs while they are away: a scheduled task. That is where the rhythm is most at risk, because nobody watches the middle. Here is how the two AI vendors this book covers support each part of the rhythm in a scheduled task today. Anthropic makes Claude, and OpenAI makes ChatGPT. Open-source tools can do the same on computers you run yourself.

Anthropic, as verified 15 October 2026. Scheduled tasks are on all paid plans: Pro, Max, Ultra, Team and Enterprise, as of 31 October 2026.¹ You set one up on the Scheduled page, or by describing it in a conversation and confirming. Claude saves your instructions, and runs the task hourly, daily, on weekdays, weekly or on demand, with your connected tools, skills and plugins. Tasks run remotely, so they run even when your computer is asleep.² As of 3 November 2026, two Anthropic pages differ on a task that needs files or apps on your computer. The scheduling page says such a task runs only on your computer.³ The page on web, desktop and mobile says new scheduled tasks run in the cloud and reach your files only while the Claude Desktop app is open.⁴ So check where your task says it runs.

• First 10: when you set a task up by hand, you choose its approval mode: Manually approve, Automatically approve or Skip all approvals. Anthropic's steps for scheduling from a conversation do not mention the setting, so check it on the Scheduled page after you create the task.²
• Middle 80: in Automatically approve, Claude checks each action for safety before it runs it. In Skip all approvals, nothing checks the task's actions. In every mode, Claude asks before it permanently deletes files. Anthropic advises starting with low-risk tasks, such as summaries, and advises against scheduling tasks that send messages, make purchases or take other actions that are hard to undo.⁵
• Final 10: each run is its own session. You review upcoming and past runs on the Scheduled page, where you can also pause, edit or run a task,² and Anthropic advises reviewing the output after each run.⁵

OpenAI, as verified 15 October 2026. Scheduled tasks in ChatGPT run once, on a schedule, or as monitoring tasks that watch for a change and notify you. They are on every plan, with limits: Free users can run a task at most once a day, and each plan limits how many tasks can be active at once. Event-triggered tasks, which respond to activity in Gmail, Slack or GitHub, are not available on Free or Go. On Enterprise and Edu, an admin must switch them on first.⁶

• First 10: you write the instructions. For an event-triggered task, you review its Trigger, Condition and Prompt before you rely on it. In managed workspaces, admins may control which apps and actions are allowed.
• Middle 80: connected apps read or act within the permissions you granted. An action that sends a message or changes data in another app may need approval. If it does, the task pauses until you review it.
• Final 10: you review results on the Scheduled page. If a task stops responding, OpenAI says to check there for a paused task or a pending approval, and to confirm the app connection is still authorized.⁶

Open source, as verified 23 October 2026. Two of the book's open-source projects run work on a schedule. Automations are OpenClaw's scheduler: one-time jobs, cron schedules, webhooks and Gmail triggers. Each wakes the agent at the set time and can send the result to a chat app.⁷ In Omnara, a cron trigger starts a new agent on a schedule, or sends input to one that is running, in the time zone you choose.⁸

• First 10: you write the job's instructions, and you set each tool yourself. In Omnara, each tool is set to always allow, always ask or always deny. Tools from connected MCP servers start at ask, and Omnara's own tools start at allow.⁹
• Middle 80: a tool set to ask stops the agent until a person approves, in Omnara's dashboard or in Slack or Discord.¹⁰ OpenClaw's tools run on the computer itself unless you turn its sandbox on, and the sandbox is off by default.¹¹
• Final 10: Omnara saves every step of every agent, so you can read what each run did.¹²

Running it safely: run a scheduled assistant on a dedicated machine, VM or container, with its own credentials,¹³ and never open it to the public internet.¹⁴

The difference that changes a decision. The two AI vendors ask for approval in different ways. On Claude, approval is one setting per task: you choose the approval mode once, and it covers every action the task takes. On ChatGPT, approval applies to kinds of action, mainly sending messages and changing data in other apps, together with your app permissions and your admin's rules. With open-source tools, you set the limits tool by tool, and the starting settings matter: some tools start at allow, and OpenClaw's sandbox starts off. Whichever you use, an action waiting for approval is your approval queue (see Lesson 3.3). A task stuck waiting for approval produces no results, so it looks exactly like a quiet night. That is why "silence is not success" (see Lesson 3.4) matters most here.

So Brightline would set up its monthly reconciliation, a scheduled task that must never change the register, differently on each.

• On Claude: the approval mode covers the whole task. Manually approve can make a person approve every step, including harmless ones, so Dave ends up running the task step by step himself (see Lesson 3.3). That is why the better design is a task that only reads and reports, with Manually approve as an extra layer of safety. Any change to the register is made later, in a separate step Dave reviews.
• On ChatGPT: the task can read freely, and it pauses when an action needs approval under the permissions and rules that apply. So Dave's review must include checking the Scheduled page for paused tasks, not just reading the reports that arrived.
• On an open-source runtime such as Omnara: set every tool that could change the register to always deny, and every tool that writes to the vendor to always ask. Then Dave's review includes the saved steps of each run, not just its report.

On all three, the strongest limit is simple: the task cannot use access it was never given. A task with read-only access to the register cannot change it through that connection, as long as no other connected tool can write to it.

Figure 3.13. Three strengths of limit. Removing every access path makes an action impossible. An approval step makes a person decide. A rule that needs judgment is written in the instructions and checked in the final 10. Its measurable parts also get automated checks.