AI Agents Run While I Sleep /loop (Claude Code Scheduled Tasks)

Key Takeaways

Anthropic just shipped a scheduled task for Claude Code and Twitter is calling it cron jobs for AI and some people are saying Claude Code just killed open clone. So, what's going on? I have set up a three agents in my Obsidian vault and I have those sub agents and I have interval when they're going to fire and let's take example here. That's my accountability check. It's asking me about okay, so every 30 minutes I'm on track on the goals. That's one example. Another one is executive coach and the final one is a source monitor. I have them here in my terminal in this workspace auto all. And here's my executive coach checking on me asking me questions. When did you feel the most alive today? Every 3 hours and then here is our source monitor which is getting information from Reddit threads and my YouTube home feed every 8 hours and another agent is accountability agent which asks me about okay, so what are you doing? Actually, I've been procrastinating on that. >> [laughter] >> So, I'm still looking that I am editing my slides for this video and saying okay, just just hit it. Just record it. And here we are. Let's try to understand how you can set this up those automations yourself. And that's what you're going to get by the end of this video. You're going to understand so what's the difference now with those scheduled tasks. We're going to set up the Reddit automation. We're going to set up accountability check and executive coach and also the limitations of this approach. All right, so let's start. So, how how can you set this up? To set this up, uh we're going to use this loop command in Claude Code. Let me open my Claude Code right now. I'm in my terminal. I go to the Claude Code. And by the way, this is also available on desktop version. Okay, um here's a loop. Let's say here's a loop command and here's a description. Run a prompt or a slash command on recurring interval. And here is a helper for us. What is the interval you want run and then there is a prompt. And let's say uh okay, so I'm going to get there and let's say um loop. So, every 30 minutes I want to read my day plan for today, my daily note and check with me if I am on track. And if I don't have anything in my daily plan for today, ask me what is the goal. Right now, Claude going to use this skill called loop. This is available by default. I didn't create this skill. It comes with the Claude Code. And now it's parsing that. Okay, every 30 minutes so interval is 30 minutes and then it's scheduled a cron job for us, basically a script which is going to run inside of this session with all of the context which we have here. And then it's going to fire every 30 minutes and it's going to read my day plan and day note and check if I'm on track. Let's actually test it. I wanted just fire this now to see how it's going to behave. So, let's let's try to fire this. Okay, so now it's going to fire it manually. So, it's going to go and explore what are my daily notes for today. So, the daily note is not on this path. So, it figured this out. Oh, are you on track for the demo day? So, I had like um two plans for today. Um and it's asking me about okay, so what's happening, right? So, I want to show how it works. So, let's actually create a a new daily note now in the daily folder and open it in my Obsidian in a new tab. And then let's say my goal for today is to publish this video and just you know, relax a bit at the end of the day. It's Saturday. Okay, so here is my daily note template which has been created and it's quite extensive. Um And let's say here is my today goal which Claude wrote. So, I want to try to see how how you're going to behave. So, try to skate like run this and uh I want to see how how it's going to look like. And here's a simulation how it's going to look like in 30 minutes. Um let's say okay, publish um scheduled task video then relax. That's a goal and then it's asking about okay, am I on track? Um where are you in the process and so on. And it's going to run it within this session. Now, and this is what makes this all powerful that we have this agent which has access to all of our skills to all of our MCP tools and connections to all of our context within Obsidian. It can reason. It can understand. It can debug and adapt at the next run. And here is a point that right? It's um running with all of the context. Let's say I have those automations set up right here. I can see what's going on here and the agent remembers what was the previous interactions before and it has all the context. And another thing which excites me that I can actually go here and if I something if I don't like something, I can just tell okay, so let's change the skill and make it in a bit different way and that's how I went about this source monitor. Um that's a Reddit and YouTube check. Let me show you what it's doing for me. So, I have this a daily note and we agreed with Claude. So, we're going to run this automation. So, I have this automations right here. It's called um source monitor. I can go there. And it tells Claude okay, so check my sources for the relevant content about Claude Code agents Obsidian. I'll scan those those Reddits using the Reddit red skill. And then also check my YouTube home feed. Okay, so and do that for each source extract uh the most interesting items and then try to notify me once we are done. And the way we did it on the first iteration this is what Claude produced. I produced this um sources with our Reddit threads. I can click there. So, it opens a Reddit thread. But I didn't like that it actually didn't read the comments and understand what what's happening there. And here is just um the YouTube videos with no links, with no summary. So, this one I didn't like and then we went about doing another iteration within the same session. And here is uh iteration two. Here's actually okay. Actually, Claude read all of the comments and extracted, you know, what is community thinks about that. And here it also for the Obsidian subreddit all of those uh different posts with activity. Now, I can go there and click and engage with the material which matters to me. So, I don't have to scroll Reddit or YouTube on my own. My agents can just okay go ahead and extract what only matters to me and what would be interesting for me. So, I don't have to go to those YouTube home feed to Reddit. Now, just um huge unlock if you can see. Now, I'm also interested in the immigration. I can check okay, so what's going on with immigration to Canada? Um there's some problems. And also for the YouTube uh this is like amazing. Just imagine that okay Claude went um and searched for the recent videos about, you know, Claude Code and that's interview with Boris Cherny. And here is actual video. So it went about getting the transcript and giving me a summary so what matters and here are some take takeaways from the actual source. You know, the Boris runs five terminal tabs um and so on. And here is some analysis. Yeah, and you can just make it your own. You can have make your own daily digest without going to those, you know, different websites. Just huge unlock. And that was a prompt which we did with Claude. That's it. I told okay, use my Reddit skill and use my YouTube skill to accomplish this. And that's how it looks like in terminal. Let me just like okay. Let me just go about um going full screen right here. So, I like it more. So, that's a source monitor. So, what we kind of did is okay, um we set up this automation and then Claude went about exploring what's going on. It appended to my daily note the first results and then it told that I want to have exact links. I want to have uh you know, uh add this uh subreddit into my monitoring and then you run it again. So, this way you can like um if you're not satisfied with the results, you can just okay, tell let's change that and let it rerun. And I can trigger this job um whenever I want. I can tell okay, so let's do this right now. So let's trigger this now. Okay, it reads this source monitor. So, it actually uh updated the skill right now because I asked it to to update before but it didn't and now it actually updates it and it reruns it right now. So, in this way you can see that if you don't like something, you can change it right away and and rerun it right away. So, here is just let's start this bug out agent. Yeah, so that's been uh this part uh for the source monitor skill. Yeah, and by the way, if you're interested in those skills, so you can grab them here. So, link in the description. So, it has the source monitor for scanning Reddit and YouTube and save it into your wall. It has to have this accountability check, which I'm going to show next and executive coach. And it's a skill pattern which you can apply for different other areas which are interested to you. The next one is accountability check. So, in this one is um the one that helps me to stay on track. Every 30 minutes, it fires this automation to see if I'm on track, if I'm on track with what I plan to do uh throughout my day. Typically, what happens I start working on something and then okay, one like two hours later I am I don't know what's happening and how I ended here in the current situation. Uh this is very very helpful um for people who might have, you know, like multitasking and they're quickly deviating away from from the goal. That's how it looks like. Um that's the accountability check. One, it's um actually uh quite cool. Uh so, it reminds me that okay, you have a limited time in your life and every 30 minutes matters. So, I'm checking with me so, what am I doing right now? So, remind today plan for my daily note. Yeah, it asks a hard question, you know? Um isn't that thing that the matters the most right now or how many like 30 minutes work do you have left today? And then in the end, it notifies me. Uh So, you can see that like I think here is a terminal um right here. That's accountability check. So, you can see that it was um actually firing firing those jobs uh at at 12 12:00 p.m. today. So, I was not answering because I was busy. And now we can try to see um let's say okay, go check on me. So, we can try try to see how how how it's going to look like. So, right now it checks what's the current date right now. So, then it fires this notification. It should appear here, I believe. Um okay, so no notifications. Yeah, the notification didn't fire, but now uh >> [laughter] >> it's telling me you're reading the accountability automation file instead of recording. Um Yes, so it's prompting me to to record. Um I'm recording. Go chill. Yeah, so you can see how it can be very very useful so to stay on track. So, this is especially useful if you're seated and think for 15 minutes at the start of your day what you want to accomplish and then you fire this agent and tell okay, so let's go, let's stay on track. So, that was accountability check. And I also have this um executive coach and I use this Dan Coyle framework for six interrupts throughout my day. So, that's the resource. Um So, let let let me let me show you um how it looks like. So, for this executive coach, it is basically runs every three hours. It tries to ask me like these questions to interrupt the patterns which I live throughout my life. Um and those are questions really really hard. So, am I moving towards the life ahead or the life I want? And those questions are really refreshing if you're throughout your day and you're you know, just um into the task and you try to think of is it really like matters? Like the most important thing to do right now? So, I would say there's overlap between this 30 minutes uh check-in and this executive coach. I think the scope is a bit different. I think this executive coach is something like which enables this like broad broad perspective on what's going on. So, I highly recommend this article from Dan Coyle. So, I think it's was right here. Um Yeah, so how to fix your life in in one day. So, it's it's a good read. So, and I basically extracted like those questions uh from this framework. Yeah, so here are the questions. Um So, regarding the throughout the day um this interrupting autopilot. And those questions are supposed to interrupt the patterns and they should be, you know, like just random. And here like those questions which I just like imported into my Obsidian and set up this agent for me. And then at the end of the day, you perform this uh insights synthesis. Um and you kind of reflect okay, so what's happening? So, then regarding the notification, so I think it's very like useful to add some observability. So, when do I need to check up on my agent? And essentially, this is done by those um notifications which you can set up. And then I tell uh when I have this automation called executive coach I tell it to run this this command to display to notify me basically that uh this session requires some action with executive coach. So, I think let's see if it's going to work right now. Yeah, I think it doesn't work because I am do not disturb mode. But that's the way you set up those uh notifications. So, also one thing which was huge and long for me is to um set up this terminal. So, I have this like terminal with uh with a bunch of um you know, tabs and it's has those workspaces. If I try to work on something else, so I create a new workspace. And then I can start Cloud Code sessions here. And it's uh it's been very helpful to keep those workspaces organized. And that's uh the way I manage my automations. So, I have this all um workspace where I have terminals running there. So, they're all together. So, I can clearly see what's happening. The terminal which I'm using is called uh Cmax. It can actually >> [gasps] >> do a lot of stuff. So, if you connect Cloud Code to this Cmax, if you can tell Cloud how to like use this terminal, it can actually do uh some quite good stuff for you. Uh let me let's say I have this orchestrator workspace and I want to show how you can set up your automation just in in one command. I want to set up my automations uh using Cmax skill. I want to create a new workspace. Um so, this is for demo purposes. I already have a workspace for my automations. Just call it in a different way. So, then I want to spawn all of the three automations in a new workspace. And then rename this workspace as well as all the terminals. Now, what's going to happen right now is that I have this um spawn workspace command which can create those Cloud Code sessions with those automations running. Now, it checks on which are the workspaces which are available right now. It checks all the smart automations. And right now it's going to set them up in a new workspace. So, it reads this automation skill right now. Okay, and here is a plan. Spawn three different workspaces. So, we actually want a one workspace and spawn uh terminals inside of this workspace with those um um automations. Okay, so it figures out that now it created a new workspace. So, that's workspace 11 right now. So, and then it creates those uh terminals for me. Right now, they're ready. So, and then it's going to run Cloud Code inside of them. So, here is uh Cloud Code which is running one by one. So, you can see it's it's working. It's sending those commands to set up a works uh those workspaces. And then here it also starts um this executive coach. And then here it's I think it's just starting. So, this one it was accountability check and here is a source uh monitor for Reddit. And then it renamed all the terminals. Uh that's accountability, source monitor, executive check. So, in this way I can just spawn my automations um and there you go. So, it's going to do all of this work for me so I don't have to do it manually. It renamed the workspaces so I can have a clear dashboard to see all of my different automations. And I just like stop stop them because I don't want to do those. They're already running for me in this tab. Yeah, so that's been uh like how you can manage those agents there through a terminal. And then also, we can control those automations through the from our phone or from a desktop app. So, here is a um Cloud Desktop app and I can actually see here all of my sessions appearing right here. And that's the one which you just run. And here we just schedule this job and I can access to those sessions from the a desktop app as well as the web app and my map my mobile phone. And that's a remote control. And the key here is to turn on in the settings, uh you can go to Cloud Code and uh say remote control and then you enable remote control for for all of the sessions. Uh each session which is spawned inside the terminal, it becomes uh available for remote control and you can connect to it through your phone or through the desktop app. It's in the same thing. Now and also that's the point that okay, so what's the difference between open claw, right? So open claw can actually do these tasks um but so what we have here is we have like full vault context on every run and you can execute skills on a different intervals and I feel it was much easier for me to manage what's happening because I was trying to set up those cron jobs inside of open claw and it was just not great experience um everything was just uh falling apart. So right now I kind of have easier way to debug my automations and my agents. And also yeah, you can use this command because caffeinate uh you can set up it in a such way that um your MacBook is going to be alive when the lid is closed. So this way you can run it um even if it's closed. Yeah, and I think that um here is the main difference. Here is the main difference um how I think about that. And this is the concept of uh this self-healing agent that um so let's say we have some automations like on this example of a daily digest, right? Where we just parsed the resources from YouTube, from Reddit. We then we run it and then we didn't like something or there was some error. So and then what we did is we provide a feedback, right? Our agent reads the logs of what went wrong. And then you can based on this feedback and based on those logs agent goes and edits the code or our skill files and then the next run is going to produce a better result. This way you can have this continuous loop where you don't actually try to design a perfect automation perfect agent from like one go. It's continuously evolving and you tune it in how you want the agent to behave and you have this complete flexibility in terms of what the agent are capable and that makes this very very flexible and just exciting, you know? And here are some limitations which I found. Here are some constraints that so those automations should be those automations only work within a session. And by session what I mean is let's say this terminal has the cloudcode running and that's a session. That's the one session of the cloudcode. And within this cloudcode if the session is alive it's going to work. Now there's another limitation that uh those sessions have a three-day expiry date. So they they auto expire and um to avoid this I also set up this um in automation base a task renewer like every two days to renew automation. So they um get automatically rescheduled um and avoiding this uh three-day limit. So it's going to delete the old cron job and then create a new cron job. Also the machine must be awake. So if your machine is not running then okay, so it's not going to run, right? And then there might be some issues with the token cost. We need to manage uh our how much is actually spent on that. So if you're going to have a lot of tasks, you can see that okay, so it might be not that manageable to to have all of them running. So here we need here we need to be mindful. And here uh here are some workarounds about that. So if you if you want to make um cloudcode session persistent persistent uh you can run it inside of the uh tmux and the tmux is a terminal multiplexer um or you can just uh set up those sessions you know, on a MacBook so on a Mac mini uh which is always on and then it runs those automations for you. And this solves this like those two issues for the session only and for the machine being awake. And I guess for this one for the three-days expiry we have this like workaround that automation which um deletes old jobs and creates new ones. And then in terms of token cost so we're going to see so how how it actually is it like manageable or not. Um yes and that's been quite quite exciting for me to record this video. I I'm very very keen on exploring this further because it seems to be very very powerful and I really love this concept that you have automation with full context with all the skills, with all of the cloud.md rules, with all of your tools which makes it very very different from what I've seen before. Now actually so if you're using cloud desktop you can also set up those tasks right there and they're in the code code code section and they're in the schedule section. So and here you can create a new task um and here is the name of the task, description and then you can select folder. You can just put your Obsidian vault um and then there is a frequency like manual, hourly, weekly and kind of similar interface. And I just found that it's it was much easier to do it from the term terminal because I don't have to fill out any forms. I can just tell okay, every 30 minutes let's do accountability and then it creates this for me. It's much easier in terminal to manage all of these automations. Yeah, and if you want to grab the skills, there's a link in the description. It contains source monitor automation, accountability check and executive coach. So you can just set this up in just one go. So link in the description and thank you for stopping by. It was been amazing to see you and I guess until the next one.