Why Claude Routines Break Down in Production
📄 Zusammenfassung
Dieses Video von "Why Claude Routines Break Down in Production" enthaelt keine Beschreibung und kein Transkript. Bitte das Video direkt auf YouTube aufrufen fuer mehr Informationen.
📝 Transkript
Recently, Anthropic announced routines inside of Claude Code. The idea is you can now automate Claude Code with prompts to run based on a schedules or events. It sounds great in practice, but is it missing a key feature? Well, we're going to have a look at it today to see if Claude Code's routines is enough to replace orchestration. Routines is simple. The idea is that the prompts that you're already using, you can now set them up as a routine. They can run on a schedule or based on an event. And so this now means you can have Claude Code running for you around the clock. No longer you're going to be hitting that 5-hour limit. Claude can run at all times of the day for you. But the question is, is this actually the best way to automate with Claude? Or is there a better way to do it? Now to find that out, we better look at what routines offers and what kind of things you might want to automate with them. Now, routines can be accessed directly inside of Claude Code by going to routines in the top left. And when you get here, you'll get the option to create a new routine, whether that's locally or remotely. Now, I'm going to set this up locally, but as you can see here, it gives us a few options. We can give it a name, a description, and then the prompt that you would normally put inside of Claude Code, but now you can have it run automatically. Then in the local example, it lets me decide if I want to run it manually, hourly, daily, weekly, or on weekdays. And when I select this, I can select the cadence of when it runs or what days of the week it runs weekly and so on. So, there's a little bit of options there, but it's fairly basic. Now the example they're giving us here is just looking at commits in the last 24 hours and summarizing what changed. These are the sort of things that you might commonly see with AI automations. Summarizing information so that when you are available, you already know what to do. Kind of kind of telling you what you need to do next. If you could start your day with the cut to do's in the priority order based on what AI has gone and looked at, that would be pretty helpful. But beyond that, it needs a little bit of human intervention, and that's kind of where this starts to fall apart. Something else that's definitely worth talking about as well is thinking about non-deterministic workflows as well as deterministic. Now, we already know that AI workflows and AI agents are non-deterministic, which means you might not get the same result every single day. Now, for summarizing commits, doesn't really matter if it gets it a little bit wrong. You know, you can probably tell if it's not quite right. But for some tasks, it's really important that the repeatability is there, which is where you would want to have that deterministic part. But this doesn't enable you to have a combination of the two. So, is there a way where you can have the benefits of an AI model, but also the repeatability and the reliability of a deterministic one? Now, if you do set these up remotely, this is where you can set it up to be event-driven, so with APIs and GitHub events as well. And you can also set up connectors. So, this is sort of you can almost think the deterministic part of what we were just talking about, but does this give you everything you need? So, I have no connectors set up at the moment, but you can see I connect this to my Gmail or whatever if I wanted to send an email, for example. So, this is a great model if you want to be able to just summarize information quickly with tools that Claude Code already has access to. But what do you do if you want to take this a step further and you want to have a little bit of human intervention involved just to make sure that it's not going to have random results? Now, that is where using a tool like Kestra to be able to do your AI orchestration or AI automations can be really powerful. Now, there's a few differences here to using Claude routines. Now, in this particular example, if we have a look at the topology view, I have three AI tasks that are running in parallel. The idea is they are going to take a little bit of information and they're going to clean it up so that it's a little bit more ready to go on the website. What this workflow does is it's going to take a new YouTube video, strip the information from it, and then add it to our website in a YAML file. So, there's a few different steps here that we need to do. Now, this is where a non-deterministic and a deterministic workflow having both bits together is really powerful because being able to generate what the title and description should look like, I want AI to do that for me and write it in the sort of tone that I would expect. The thing is that's not repeatable. Every workflow is going to have slightly different results. So, being able to pass that off to an AI agent and have that write it for me is very, very powerful. But then once we have that information, being able to then use the deterministic part of the workflow to create the file, add it to the Git repository, and then make a pull request, all of that is repeatable and nothing about that changes. As long as we're using the same information from these three tasks and passing it in dynamically to these ones, this bit is going to repeat itself every single time. Now, this also is where token cost comes into it because I'm only spending tokens generating those three bits of information at the top, the category, the title, and the description. Everything else isn't costing me any tokens, whereas using Claude Code, some of those deterministic steps, which the AI model can do in a non-deterministic way, they're going to cost you tokens as every single time, even if it's getting the same results. So, that's where having a combination of the two can be very, very powerful. Now, the last thing to mention is human in the loop intervention as well. Now, this workflow does not automatically add them to the website. It goes for a pull request review, which means that someone else has to review it before it goes live on our website. And don't get me wrong, sometimes the AI agent doesn't always suggest the description or the title that I would put. So, being able to have a human verify that it is in fact correct before it goes live is very, very important. And this is something again where Claude routines starts to fall flat as you sort of build up your automations. So, for basic automations, it's really easy to set up and it's really powerful. But as soon as you want a little bit more control or you want to be able to be a little bit more conscious about your token spend, that's where being able to break it into a full workflow with multiple steps is very, very powerful. Now, for example, here's the pull request added from my workflow. And as you can see here, I had my technical writer AJ come and approve the pull request before it got merged so that we can clarify that in fact this is exactly what we wanted it to say on the website before it gets added. Now, there's another thing I want to talk about as well, which is not just using a combination of deterministic and non-deterministic inside of your workflow, but being able to kind of give AI a little bit more structure on what it should output. Now, obviously you can tell it as part of the prompt what it should return, only return these things, keep it concise, whatever. But that's still going to potentially give you a variety of results. Now, a lot of the AI models, especially when you're using their API, have a thing called structured output, which is where you can give it a JSON schema of how you want it to return the data. Now, here's a workflow I built previously that checks the weather for a sport, and based on the weather, it should basically say yes or no. We can then pass that yes or no to an if statement that will then be able to go away and do some deterministic tasks, such as sending a Slack message and adding a calendar event. Now, to make this really easy, I've given it this JSON schema here where I've asked it to give me three different values, what type I expect them to be, which makes it really easy. So, for example, this go sailing one here, the type is a Boolean, which makes it very easy to throw it into this condition here because I know it's always going to be a Boolean. It's going to be true or false. Now, again, non-deterministic, it could get it wrong. It could say false when it means true, but it's not going to say something else that's then going to cause the rest of my workflow to break. So, being able to use these JSON schemas is really powerful. And my understanding is all of the major AI companies do support this in their APIs. So, hopefully that's given you a bit of an insight into whether you should use Claude routines or if you should build more advanced workflows using a tool like Kestra. Both have their place, but as you start to build out those automations further, you definitely want to have a look at an orchestrator, which could both help you manage costs, but also give you a little bit more control over what tasks are non-deterministic, but what tasks are deterministic. Hopefully you found this video helpful, and if you enjoyed it, make sure to give it a thumbs up, and make sure to subscribe to the Kestra YouTube channel if you'd love to see more videos like this.
📺 Ähnliche Videos
🔮 MiroFish simuliert die Zukunft mit tausenden Agents | Alles was du wissen musst!
Christoph Magnussen · 2026-05-11🇩🇪 DE
Ich habe das fortschrittlichste KI-Tool der aktuellen Zeit entdeckt… Das ist Hermes Agent 🔥
Der KI-Doktor · 2026-05-09🇩🇪 DE
Diese OpenClaw MasterClass Wird Deine Arbeitsweise Für Immer Verändern
Der KI-Doktor · 2026-05-08🇩🇪 DE
Paperclip Is Insane | Full Tutorial
Ferdy․com | Ferdy Korpershoek · 2026-05-06🇬🇧 EN