Claude Code Agent Loop Can't Replace OpenClaw (Here's Why)

BBetter Stack
Computing/SoftwareSmall Business/StartupsInternet Technology

Transcript

00:00:00The Anthropic team have released a loop skill for Claude Code to run prompts at set intervals
00:00:04and people on the internet are going crazy, hooking it up to Discord or Telegram to communicate
00:00:10with their agent just like they would with OpenClaw, and some are even using it for Ralph
00:00:14Loops.
00:00:15But they're using it all wrong because the loop skill was not designed for that purpose.
00:00:19I mean, there is a reason they didn't call it schedule and call it loop.
00:00:23Hit subscribe and let's find out why.
00:00:29It's no secret that Claude is being used to write 100% of Claude Code because the team
00:00:34have been shipping features like crazy, but usually the hype level matches the usefulness
00:00:39of the feature.
00:00:40So Claude remote control, lots of hype because it's very useful.
00:00:43Async hooks less hype because they're less useful.
00:00:47But this is the first feature that has more hype around it than the usefulness of the actual
00:00:53feature.
00:00:54Before you get angry, let me explain.
00:00:56So here is the latest version of Claude Code running Sonnet, which I'm going to change
00:01:01to Haiku since the prompts I'm going to use don't require such a sophisticated model.
00:01:06So here I just have the loop, which takes an interval in minutes, hours or days.
00:01:10Note the minimum granularity is in minutes.
00:01:13So even though you can write seconds, Claude doesn't support them yet.
00:01:16And then it also takes a prompt, so I'm going to write, say hello.
00:01:20And if I hit enter, it uses the cron create skill to create a scheduled task.
00:01:25You can also set a specific time so I could do loop and here it said hello very quickly.
00:01:30I'll write a prompt of say hello again at 6.30 PM.
00:01:35And here Claude uses the local time for the machine you've installed Claude code on and
00:01:39not UTC.
00:01:40So keep that in mind.
00:01:41I'll set this to a daily recurring task and that's pretty much it.
00:01:45The prompt can contain anything you want.
00:01:47So skills like I could use this tweet skill to write me a tweet along the lines of Claude
00:01:52has an awesome new loop skill and I'll set that to run every three minutes.
00:01:56The prompt can also read a file, run MCP tools.
00:01:59Basically anything you can do with Claude code you can add to your prompt.
00:02:02And I could list the current cron jobs running with this prompt, which uses the cron list
00:02:07tool to list out the jobs.
00:02:09And I could delete a job by saying what the job does or by using the ID.
00:02:13There's also a jitter on the time of cron jobs to prevent multiple jobs at the same time hitting
00:02:17the anthropic API, which means an hourly job can fire anywhere between zero and six minutes
00:02:23on the hour.
00:02:24Now, yes, this is a really cool feature and people are saying it has replaced Open Claude
00:02:28for them because you can set it on an interval to pull messages from WhatsApp or Telegram
00:02:33and respond to them like a real AI assistant, which is what people were using Open Claude
00:02:37for doing.
00:02:38But there are two major issues with the loop feature that prevent it from being an Open
00:02:43Claude killer.
00:02:44First of all, the cron tasks auto expire after three days to prevent you from having long
00:02:49running tasks that you might have forgotten about.
00:02:51This could of course be really annoying if you've hooked the loop up to listen for telegram
00:02:55messages and go to it on the fourth day to find out it doesn't work.
00:02:59Now there is a way to run jobs forever inside Claude code without a plugin, but we'll talk
00:03:04about that later on in the video.
00:03:06The second issue is that cron tasks are stored in session memory and are not written to disk.
00:03:12So if I close this Claude code session, I'm going to clear my terminal and create a new
00:03:16one.
00:03:17Then if I ask it to list my scheduled tasks, you'll see that nothing has been scheduled,
00:03:21even though I scheduled two tasks in the previous session.
00:03:25This means if you wanted to close the session to update Claude or if you wanted to use the
00:03:29loop feature for a Ralph loop, it isn't really a good idea since you can't properly reset
00:03:34the session and have to rely on compaction to continue long running tasks.
00:03:39So what should you be using this feature for?
00:03:41Well, whenever you have to rerun a task manually in the same session, because something has
00:03:46changed, then you should be using a loop for that.
00:03:49For example, if you're checking the last 50 lines of a continuously updating log file
00:03:53for errors, seeing how many items are left in a job queue or checking if you have any
00:03:57new issues in your project.
00:03:59Now I'm sure the team will add more options to loop in the future, like setting a specific
00:04:03model for a task or changing the expiry time.
00:04:06But if you want a task to run forever in Claude code, you can kind of already technically do
00:04:11that without any plugins.
00:04:13You can do that in Claude for desktop by opening the sidebar if it isn't already and clicking
00:04:17on scheduled in the menu.
00:04:19Now if you don't see this scheduled option, you may have to update Claude for desktop.
00:04:23And here we can create a task, giving it a name, description and prompt.
00:04:27We can even change the model, set permissions and set frequency down here.
00:04:32This is exactly the same as if we were to write schedule inside the prompt section of Claude
00:04:36code desktop, but what's really interesting is that inside cowork, there is also a scheduled
00:04:42option.
00:04:43Now this is different from the scheduled option in Claude code, not only because the modal
00:04:47looks different and we have to click on more options to change the default model, but also
00:04:52because Claude cowork runs inside a sandboxed environment, whereas Claude code runs on your
00:04:59local machine.
00:05:00So if you schedule a task in Claude code, that's going to use your local file system.
00:05:05Whereas for Claude cowork, it's going to do it in a sandboxed environment, which is important
00:05:10so you know where to put your scheduled tasks.
00:05:12Now the benefit of adding a schedule task inside Claude desktop is of course the task will run
00:05:16forever as long as the computer is switched on and the Claude desktop app is opened.
00:05:22But if you're like me and would rather use Claude code in the terminal to schedule tasks,
00:05:26then you can use this plugin by Kenneth, which does look promising and I may try it out in
00:05:31the future if there's enough interest.
00:05:33Speaking of other promising things I haven't yet tried out, open AI released GPT 5.4 last
00:05:39week and people seem to be excited about it.
00:05:41If you want to know more, check out this video from James who goes through all the details.

Key Takeaway

While the Claude Code loop skill is powerful for short-term session tasks, it cannot replace persistent agents like OpenClaw due to its 3-day expiry and lack of disk-based memory.

Highlights

Claude Code's new 'loop' skill is designed for recurring tasks within a session but is being widely misunderstood by the community.

The loop skill uses local machine time and includes a 'jitter' of 0-6 minutes to prevent API overload.

Major limitations include a 3-day automatic expiry for tasks and the loss of all scheduled tasks when a session is closed.

The feature is best suited for monitoring log files, job queues, or project issues within an active development session.

Permanent scheduling can be achieved through Claude for Desktop's 'scheduled' menu rather than the transient loop skill.

Claude Co-work offers a similar scheduling feature but operates in a sandboxed environment compared to the local execution of Claude Code.

Timeline

Introduction to the Loop Skill and Community Misconceptions

The speaker introduces the newly released 'loop' skill for Claude Code, which allows users to run prompts at set intervals. Many users are mistakenly attempting to use it as a persistent assistant by connecting it to Discord or Telegram. The narrator emphasizes that the feature was specifically named 'loop' rather than 'schedule' for a technical reason. He notes that while Claude Code is being developed rapidly, this is the first feature where the public hype significantly outweighs its actual utility. This section sets the stage for a deeper dive into the technical constraints that define the tool's true purpose.

Technical Demonstration and Configuration of Loops

The video demonstrates how to initiate a loop using Haiku for simpler tasks to save on costs. Users can set intervals in minutes, hours, or days, though the speaker clarifies that sub-minute granularity is not yet supported. The loop skill utilizes the 'cron create' tool and relies on the local machine's time zone rather than UTC. Practical examples shown include automated tweeting every few minutes and reading local files or MCP tools. A critical detail mentioned is the 'jitter' mechanism, which can delay tasks by up to six minutes to avoid hitting Anthropic's API limits simultaneously.

The Two Major Limitations: Expiry and Memory

The speaker outlines why the loop skill fails as a replacement for long-term agents like OpenClaw. First, all cron tasks automatically expire after three days to prevent forgotten, long-running processes from consuming resources. Second, these tasks are stored only in volatile session memory and are not written to the disk. This means closing a terminal or restarting a session completely wipes all scheduled tasks, making it unreliable for persistent automation. These constraints mean users cannot easily reset sessions or rely on the tool for permanent 'Ralph loops' without constant manual intervention.

Ideal Use Cases and Desktop Alternatives

This section explains that the loop skill is intended for repetitive tasks within a single, active working session. Examples include monitoring the last 50 lines of a log file for errors or checking a project's issue queue for updates. For those needing permanent tasks, the speaker suggests using the 'scheduled' option in the Claude Desktop app. He distinguishes between Claude Desktop, which accesses local files, and Claude Co-work, which operates in a sandboxed environment. Understanding the difference in environments is crucial for users to decide where to place their automated tasks safely.

Third-Party Plugins and Closing Thoughts

The video concludes by mentioning a promising third-party plugin by Kenneth that might solve the persistence issues within the terminal. The speaker expresses interest in testing this plugin if viewers show enough curiosity in the comments. He briefly touches upon the release of GPT 5.4 by OpenAI and points viewers toward other resources for those specific details. The final message reinforces that while the native loop skill is a 'cool feature,' it requires specific workarounds or different platforms for long-term reliability. The content ends with a call to action for subscribers interested in advanced AI agent workflows.

Community Posts

View all posts