In AI agent automation, heartbeats are timer-based triggers that wake agents regardless of whether there is work to do, potentially wasting credits when queues are empty, while routines are specific scheduled tasks (like Gmail polling at 9am or content drip at noon) that only execute when there is actual work, making them more granular, intentional, and cost-effective for agent operations.
深掘り
前提条件
- データがありません。
次のステップ
- データがありません。
深掘り
Paperclip Heartbeats Are Out. Routines Are In. Before June 15th — Do It.追加:
Hey, what's up everyone? Aaron Prince.
It's been a while since since I recorded a video, but I built recent changes with Claude and you know I just felt that I had it to go on camera really quick. Record a quick video for you guys some updates.
So first of all, a lot has happened since the last update. I've become really involved with Paperclip. There's some really cool things behind the scenes going on that I want to not really cover but cover a little bit. So first of all, the Paperclip desktop app is the thing now.
If you go to my repo not that one and or to my GitHub account you actually see Paperclip desktop.
Um >> [clears throat] >> Here we go and it is actually free to download. You simply go to releases.
You download the the one the Mac app that you need. This one is for Intel Macs and this one is for all you fancy guys with the new Apple silicon and it basically gives you a wrapper around Paperclip. So it doesn't really replace it. Just pulls in Paperclip like you would run it manually, but now you just click the app and it loads up the server. You don't need to go into terminal and all that good stuff. You can also use a Paperclip desktop remote to connect to an external Paperclip setup that you have an account on.
That's it for for that.
In the meantime, I have also been put in charge of Paperclip documentation.
I've actually created this this whole framework here which is powered by my Cleo KB Cleo knowledge base.
Here is at cleokb.com and this one is the Paperclip specific adjusted version which is tweaked out and then redesigned a little bit. Uh but I've actually pulled the open-source Cleo from that and published it for you guys to use.
Um but here we go through everything that happens in Paperclip, um which I keep active. I actively update every single release.
Um we're actually getting into the rhythm of of of doing that. So, so that's really cool. So, if you need any help with Paperclip, you can go to docs.paperclip.ing ing. Um and uh that actually gets you that. Uh then I have launched the Paperclip community. You can find that at paperclip.community. You can join for free.
Uh ask any questions that you might have about Paperclip, and I or someone in the community, I'm sure, will get back to you.
Um we have some really cool things lined up. We have some guides. The documentation is available there as well. Uh and the resources um that you can find some cool plugins, templates, um adapters, all sorts of cool stuff that you can use for Paperclip. So, that's what I sort of want to build the whole thing around. Um and um you know, just be of help and support to you guys. Um there is a big update coming in the next few days called companies, which is going to be available here at the top menu, where you can actually add your own companies to share with the community. You can also fork companies, and there's a whole company editor where you can visually edit your company. And then once it's done, you can export it uh and import it into Paperclip. So, this also means that if you're looking for a company template, you can come to the paperclip.community.
Uh go to companies, find a company that matches what you need. You either download that and import it into Paperclip, uh or you fork it, you edit it, and then you import it into Paperclip. Uh so, that's really useful.
Um and then I'm also working on an AI integration, of course.
Uh wouldn't be wouldn't be an AI company uh if if I didn't wouldn't do that, which interviews you, um it's basically a sort of chat interface, but very strictly, very guided, um that asks what needs to be asked. And based on that, it generates a whole bespoke company for you. That is going to be a paid service, um but for now, what's planned with the company is it's going to be minus the AI, it's still going to be free. Um so, that is that. And then, if we go to my X, uh follow me.
Uh if we go to Yeah, I've been actively tweeting about this, but I want to cover this for you guys. Uh because this is going to be a change to Paperclip or Paperclip usage as well.
>> [sighs and gasps] >> Uh Claude is playing up again, or Anthropic, whatever you want to call it.
And um I really really I want to make this very clear, they are full well within their right to do anything they want. Um that is the price that we pay as end users, uh and that we're becoming dependent on uh on these on these models.
And the way that we the way that we use them, and now, um yeah, things are just changing.
Um so, I don't want to bash it too much, but I want to make sure that you are aware of what's going on. Um the Claude - P, which is to my knowledge what Paperclip wraps around, Paperclip uses Claude.p. I believe that Dota recently integrated the agent SDK as well. Uh so, that that's for good, but um but the way that we're going to be billed for it is going to change. Um there's going to be I don't think it's here yet.
No. So, there's going to be a new bar here that's going to be um agentic usage agent SDK usage, whatever they're going to call it. And it's going to be a separate bar that's going to give us an uh a credit allotment, a credit amount uh for the amount, I believe, that we pay uh for our subscription. So, I'm personally on the $200 plan. So, I'm going to get $200 of monthly credits for minus p agent SDK usage. Once that drains out, I'm going to have to turn on extra usage if I want to continue using Claude for some, you know, specific tasks, whether that is with a rough loop or paperclip. I will update on my X for this. I will update in the community if I have to, but I believe again that paperclip is wrapping around the agent either the agent SDK or minus p.
So, this is going to affect paperclip for sure. Some some way.
I also know that if we go to a new issue, there are I believe there was a function here that I saw that I don't see now, which is interesting.
But, it allowed me to Oh, here we go. I have to select something first, and then I can see the Claude options, which allows me to go with a custom model for that specific issue.
Though, it had really went, you know, prepped for this apparently.
This was by the implemented just before this announcement. So, this is really neat, but I think that this is going to be expanded very very big into paperclip.
I think agent model routing is going to be is going to be of interest. But, I want to make sure that you are aware that this is going to happen. So, right now if you use Claude code up until what was the date? June 15th.
Programmatic usage, which is what Claude is or paperclip is using, gets its own dedicated budget instead. And you'll be able to find that in your usage page for Claude code. You can see your meter run out. And then you're going to have to have to turn on the you know, the charging the the extra usage or you're going to have to use another model. So, to sort of mitigate for this, I have already and you can find this in the documentation.
Uh just look for heartbeats and routines, which is the docs article.
Um So, docs.paperclip.ing in the top search bar you go for heartbeats.
When you first [snorts] hire agents, it's tempting to give each one a timer, which is the heartbeat.
This a couple of versions ago used to be on by default. Now for new agents, this is off. If you create if you just onboard to Paperclip, your main agent is going to have it turned off. So, there's no need to directly worry about this, but in the past, there was a heartbeat that would trigger say every 30 minutes, every hour, every It was it it's config it was configurable.
Um for every single agent or it's still is configurable by the way.
Uh so, if you go to a specific agent, my apologies for the capacity. Go to a specific agent in your company. You click on configuration. You scroll down a little bit and you're going to find the run policy and it's going to be heartbeat. You could turn on or off and um set an interval. So, run it every 1800 seconds, which I believe is half an hour.
3600 then, you know, is like an hour etc. etc. So, right now, um we have the support for wake on demand where you turn that on and um actually other agents can tag each other and collaborate and work together.
But the main replacement for this is going to be routines. It's not a replacement if you need heartbeats, it's there, but routines are what many of us are going to be you know, going to be after, which you can find here in the sidebar.
And a routine is basically an issue that's on repeat. So, for example, I use email communication and for my me and my agents, I use the GOG CLI.
Um Do I have this in the documentation? I don't have this in the documentation. Uh but I do have an article in the community, uh a guide, whatever you connect your agents can use, and the GOG CLI is in there. Here we go.
Uh workspace awareness is with GOG CLI, which is a tool by Peter Steinberg of um Open Claw fame. And um I use that for my Gmail poll. And um it emails me every single day with, you know, what's going on and what needs approval from me, and then I just reply to that email and it picks it up and starts to doing the work.
Uh on a heartbeat, right? Automatically.
It just tags the agents that have to be tagged. But obviously, this is way cheaper um because I can I have more granular control over when it runs. Um so, um you can read the article either in the community or the official Pipedream docs, but I recommend you read this article because it's going to help you move away from heartbeats. It's going to save you a lot of money. And you can assign specific routines to specific agents, which also means that you could um you could set up a routine for a specific agent, and then I'm sure I'm not sure if this is available yet, but I'm sure it will be where um is it available? Create routine.
Delivery settings. I don't see it here.
Which one is safe for me to show you guys?
This one is archive, so this is sort of Okay, yeah, I can share this.
Um Advanced delivery. Yeah, there's no setting for it yet as far as I can see, but I'm sure it will be added where um you actually configure the model based on the routine.
Which then means that depending on the routine, the task that you want to have happen on repeat, uh you can actually configure that per routine uh and defer what we can, right? So, we have to be a little bit granular about uh a music bot from now on or at least from June 15th on unless this changes miraculously unless they prefer that decision. Um, that is what you want to do. So, for example, I have a a [snorts] QA reviewer, which I believe I have already configured. Oh, I don't have it configured to Codex. Uh, that's in another company probably. But, um uh I usually use um uh Codex for um QA. So, uh the once the QA QA guy gets in the loop, it's another model looking at it. That's how I work when I do work.
Um, uh in the Claude code or in Claude, I have Claude do the the cool stuff. You Claude's like the cool guy and then Codex is like the PhD smarter uh guy uh that actually checks the the juniors guys work the the the like the surf dude's work. Uh, and then Codex says, "Well, you did this wrong and you did that wrong and I found this and I found that." And I'm like, "Okay, Codex, can you fix that for me before I push it to production?" And I have that running a loop until, you know, the work is done.
But, uh that is how I'm going, you know, going to have to be uh using it. Uh Nikki Digital Tricky Magical on Discord actually gave me the tip uh just now for something called manifest.build. I need to look into that. It's like a model router.
Um, same with the open code. So, you know, open code uh and open router are are things that we're going to have to look into probably too. And I believe I've read a few mentions from Dota um that uh that he's considering training some type of model himself. So, that is interesting as well. Um, yeah, that's that. I just wanted to tell you this. Dive in with you a little bit.
Um, so um you're going to have to we're going to have to look at what models work best, you know, for um for what purpose. Um, And yeah, eventually companies are going to be helping with that as well, like the the thing I'm building in the community.
And to basically see what works really well.
Maybe we should do some polls and see, you know, either here or in the docs or on Discord what model works best for what.
But yeah, some things are going to be changing, which is cool. We're going to have to adapt to that. And um yeah, I think that's that's it for this video. So again, go to docs.paperclip.ing.
Look for heartbeats and routines. Read this article or have your agent since your agent here have it read it.
One last tip I want to give you is that agents can only work on the routines they own. So we can't have the CEO handle and manage and operate all the agents or all the routines for all agents.
So they'll have to be instructed individually through the CEO in that case to do that. And then your heartbeats probably getting quite long quite fast.
So what you do is you ask it, "Hey, look at your own heartbeat and extract what you think can be turned into routines.
And then you have it turn it into routines and check the work's done. But that basically automates, you know, a bit of that bit of that work. These three that are live right now are based on routines.
That's it for this video. Aaron is checking out. Hope it helped. And I'll see you in the next one. Take care. Bye.
関連おすすめ
OpenHuman VS Hermes AI: Who Wins?
JulianGoldieSEO
285 views•2026-05-29
Long-Running Agents — Build an Agent That Never Forgets with Google ADK
suryakunju
142 views•2026-05-30
5 Mind Blowing Omni Uses Cases
PaulJLipsky
1K views•2026-06-02
This computer is made from real human brain cells. And you can buy it.
Talktmsmedia
3K views•2026-05-28
BREAKING: Microsoft’s New Image Generating Model Beat Out GPT 1.5 and Nano Banana 2
aimmediahouse
122 views•2026-06-03
I Made the Same Anime Fight Scene in Every AI Video Generator
NobleGooseAnime
295 views•2026-05-30
Nvidia Bets Big On AI PCs | New Chip To Power Windows Laptops | Technology | AI Updates | N18S
cnnnews18
3K views•2026-06-01
I Tested NEW Opus 4.8 on Four Projects (Updated LLM Leaderboard)
AICodingDaily
298 views•2026-05-29











