So What? How to Get Started with Paperclip AI, the AI Agency Software

Well, hey everyone, happy Thursday. Welcome to So What the Marketing Analytics and Insights Live Show. I'm Katie, joined by Chris. John is traveling, but George is here. Yeah, John's out at Salesforce Connections.

He is, yeah. So if you're watching this uh as of today, which I don't know, maybe you are, uh, go say hi to John. He's in Chicago. And if you miss him, then come say hi to him at uh trustinsights.ai slash contact. Today we are talking about how to get started with paperclip AI, the AI agency software.

Uh earlier this week on the podcast, which you can listen to at TrustInsights.t.ai slash TI podcast, uh, we talked about what paperclip AI is. Today, we're going to talk about setting it up and actually using it. And so at a high level, Chris, my understanding is that paperclip AI is another generative AI piece of software. But the idea with paperclip is that you can basically, we're saying you can create your uh AI agency, but you can create these, you know, versions of people who can actually talk to each other and remember the context of conversations. So, you know, we're sort of equating it to um a company that isn't siloed and actually shares conversations and resources versus the way that a lot of large language models operate today, which does kind of come across as a siloed company where it's everyone's in their lane and nobody shares information unless you like really force them in the room together to do that.

Is that at a high level, very high level, a good way to think about it? Yeah. Here's the thing about paperclip. I've because I've been thinking about this as we talked on on Monday about the podcast on the when we did the podcast. Paperclip, they bill it as the agentic control plane for autonomous AI companies, which is a lot of words.

That is a lot of words. That's a lot of buzz. It is a lot of buzz. So here's the buzz-free version that Katie, I think you'll wrap your brain around immediately. Okay.

It's Jira for AI. Oh. Got it. Yes. Uh, so for those who uh may not be familiar, describe Jira.

Uh Jira is project management software where you set up projects, you know, things, different things of people working on your company, you know, uh, and it could be any department, coding, development, UI, UX, customer service, or whatever. People open tickets and assign them to other people in other departments. Like, hey, a customer just called in, this thing's not working. You assign that to the development team, the development team routes it to a developer, the developer works on it and says, Yep, I solved this problem for this customer, and so on and so forth. So at a high level, when we talk about things like JIRA, so it is a project management software.

Project management software a lot of times is meant to be uh the the rec system of record for like tasks and deliverables and documents and the flow, Chris, that you just described in terms of like a ticket going through people. It's that ticket that allows for the collaboration between you and me and the developer and someone else, because that ticket has all of the information that we all need and we take from it what we need. And so I think that that's an interesting way to think about it, which is not the way that I would assume AI agency software uh is positioned. So, should we get into it? Like what are we doing?

Yeah, we can get into it. So we're not gonna talk about setting it up because setting it up, frankly, is like a five-hour ordeal. Oh, yeah, we're not gonna do that. Not even with the uh the Benny Hill sped up music. At a very high level, yes, at a very high level, you should run it on a VPS or a dedicated box that has to have access to a gentic software of some kind.

It can be clawed code, it can be open claw, Hermes Agent, anything. And we in fact we did a uh show on Hermes Agent, what, two weeks ago on the live stream? Um like a month ago, Chris. But if you're interested in watching that, you can go to trustinsights.ai/slash YouTube, go to the so what live stream playlist. Yeah, things move pretty quickly.

So that was about a month ago. I I've lost all track of time. Um, essentially, if you have an agentic tool that can, it's called receive a heartbeat, aka this software can ping it and say, Hey, you're there, can it can essentially prompt it? Uh, you can use it with this. And so those are the requirements to set up paperclip.

I strongly recommend, like we said with Hermes Agent, that you have it on its own little sandbox so that it is not on production systems where other important stuff lives. Please do not run this on your main computer and give it blanket access to everything. You will be sorry. And not I mean, that's in general just good best practices for software development, which to be fair, a lot of these large language models are software development. Exactly.

So what we see in our agentic control plane is um four sections. So there's the work itself, there's projects, agents, which are your virtual people, uh, and then the company, the org chart, the skills, etc. So again, if you think of it as like JIRA, JIRA would have projects or departments, right? Jira would have people in it. Jira has a ticketing system.

And then you have some like general reporting and stuff, like, okay, or what's going on in our company in our in our project management system. I mean, honestly, it's very, at least for me, and this sort of like takes me way back, it's very reminiscent of uh Microsoft Project connected with SharePoint. So, you know, the way that Microsoft Project was the most effective was if you could get all of that company information into it, so then you could choose your agents, which are really just people, to say, you know, Chris is the chief data scientist, this is his you know, hourly bill rate, these are his uh responsibilities, these are the projects they already has. So that when you're setting up a project, you have all of that background data to say, is Chris even available for the month of June? Well, he already has five other projects assigned to him, and this is what it's gonna cost us to get those done.

Can we add in a sixth one? Does he have a spare hour here or there? And does the budget of your project support it? Like I'm looking at this, I'm like, yeah, you're right. It is project management software.

Exactly. Uh, which makes it a lot easier to understand it, it does take away the the uh sexiness of all the expensive words, right? But it also means that if you know project management well, you know how to use paperclip because the bunch of stuff that's not in here. There are no you know, Gantt charts, there are no uh workflows like that, because you don't need it for AI agents. AI agents are assigned to go do things and are told to go do things and they go off and do things.

Um you don't have to say, like, does Perme's agent have time to work on this today? It that's irrelevant uh in this system as it stands today, but the caveat there have been 12 major updates in the last two weeks to this. So this is a very, very fast-moving project. Okay, I don't know. When I hear something like that, that to me is a huge red flag.

But anyway, as we said on the podcast, this is not production ready. No, no, no, it's a shiny object. It's it's a shiny object, and it's a good idea, it's a concept that you're going to start seeing more and more of. So some of what Paperclip does, for example, anthropic just built into Claude code. It's called dynamic workflows, where Claude sets up its own project management system internally to execute on a task.

What's different about that versus this is this has persistent memory. So it remembers what happened. Kind of like you would use Jira to say, Well, what have you been doing all week, Claude? Uh whereas Claude's like, I'm doing things, and then then you know you turn it on again, it's like, I have done nothing. You're like, okay, great.

Huh? It's like having a whole team of people. It really is. All right. So the first things first, we as you do with project management software, we're gonna start a project.

I'm gonna name it test. Uh it was called call test bed. You can configure it to connect to a GitHub repo if you want. If you're doing things like software development or you want to have a repository where humans can look in on things that's sort of centralized, that's a great way to do it. Um, and then if you have like deadlines, you can put them in here and so on and so forth.

So we're gonna create that first project called test. Next, we're gonna create our first employee, uh, which is just an agent. Sure. So there's a bunch of different ways to do this. Uh right now there is no CEO agent at all.

There's nothing in here at all. So this is kind of a an interface bug. Um, yeah, I was gonna say that's a weird thing. Yes, it's a weird thing. Well, if you set it up net new, you create a generic CEO agent up front, which is actually a terrible idea.

Um, because Katie, you want to talk, you know, uh re uh review what you said about that on Monday? Yeah, absolutely. So you can hear the full rant at trustinsights.ai slash ti podcast, but my concern, rant, uh, is that you know you're setting up net new. Maybe you've never been a CEO, maybe you don't have a lot of experience being a CEO. You're just kind of making it up, and or you're just grabbing maybe CEO job descriptions, which are never accurate.

Job scripts are never accurate, uh, from the internet, and what you think a CEO does. So it's not to say you can't come up with a you know good proxy for a CEO, but I do think that it's also you know a huge risk to ask someone who's maybe brand new to their career or has always worked in like a different vertical or a different kind of discipline to say, okay, first and foremost, you need to create a CEO. If you don't have a good working understanding of what a CEO does, then you could be making a lot of assumptions, you could introduce a lot of bias. I think that you should first and foremost create like an agent, but calling it a CEO to me is sort of a risky move because I feel like it introduces a lot of things that may or may not be true. Yep.

Yeah, it would be equally appropriate to call it like a project manager. Um, because again, if this is JIRA, you'd have a PMO and there's a head of a PMO, and that would be just as good for this. Yep. So we're gonna create the runtime manually. We're gonna choose Hermes agent.

You can see all the different tools that are available to you. So if you have something else, if you use open code, if you use cursor, if you use clawed code, you can you can use them in here. I'm gonna use Hermes Agent because I know it's configured, and we're gonna call this Katie, the CEO. Now, here's what we're gonna use for Katie's um we have thinking effort. The model is the default, which is minimax.

Hermes is the system. Uh Hermes is the local agent. There's the run policy. Uh, there are no company skills set up yet. Um, so I'm gonna set, I'm gonna create the agent.

And now we're gonna create a uh, oh, these these are the default skills. I'm gonna add a skill, and I'm gonna create a blank skill here. Uh and the skill we're going to call it is this should be very familiar, co-CEO. Uh and we have built that co-CEO on numerous live streams. So again, you can get that at the live stream uh playlist on our YouTube channel.

The co-CEO has been around for at least a year in our ecosystem. Exactly. So rather than me try to make up and invent what a CEO should be, we reuse what we've already done. So if we talked about on past live streams and podcasts, the five levels of AI, right? From level one, which is chat GPT, to level two, which is gems, which is what the co-CEO is, uh, to level three, which is a agent systems like Claude Code, to level four systems like Hermes Agent, which are more autonomous, to level five, so what's the paperclip idea is if you did your homework and you've been working up your the ladder, not trying to skip ahead, as we just did, we fished something out from level two, the co-CEO's uh gem or skill and said, Okay, let's make use of this.

So I'm making sure that co CEO is installed. So now Katie is is ready. Katie the uh AI CEO is ready to work. Great. The next thing great, great.

Um, and this one is the master agent that can then go and hire other agents if needed. So we're gonna do a very, very simple test. Uh, and we're gonna open our first issue. But this is again like Jira, this is exactly how uh uh work gets done. And this is the single unit of work is the issue.

So let's come up with um a simple task for this. So we're gonna create our first issue with the issue, we'll just call it um meeting notes idea. We're going to make sure it's assigned to Katie in the test project, and this is where we specify what the task is. So I'm gonna say write three plain English bullet points exam, explain what good meeting notes should include, right? Something super simple so that we can test that this thing is actually working or not.

We'll give it a critical priority, we'll make sure it is on to do, because you can see backlog. It won't, nothing will happen. Um, there's no additional stuff here to do. There's uh I could create start dates, I'm not going to create the issue. And now what we should see is Katie the CEO now has an indicator that it's live.

And on the dashboard up front, Katie the CEO is working on this very, very simple task. Now on the back end, you can't see it here, but on the back end, um, the system has started to route this, turned on Hermes agent and is burning away. But effectively, this is what's happened. This is what happens in paperclip. You can you can imagine now how you could maybe set up additional agents or have the CEO create additional agents if you know your overall project plan.

Um you could design really good project plans with multiple stages and decompose them. Um a lot of the way people are using this, I think I don't want to say it's wrong, but it's kind of wrong. Um, is just winging it and say I've seen the example uh, for example, on the paperclip uh subreddit saying, make me a new product that will earn a million dollars in recurring revenue. That would that's just the task they give in it, and then they just kind of fling it to the to the void and hope it works. I mean, I'll go ahead and say it.

That's the wrong way to do it. I'll say it. There's no probably about it. Um, no, I mean, experimentation, research and development, all of those things are valid, innovation is valid, but when you bring it into a real company where real dollars and real resources are involved and at stake, you really want to make sure you have a good solid plan. Like I think this is a good example.

Um, you know, a real issue, not issue, a real challenge that we have is that we are a small company. And so we were talking on a previous meeting about you know, doing uh promotions, uh advertising, marketing for our academy, and it's something we all know needs to get done, but because we're a small company, it always kind of falls to the wayside. So I could see saying, okay, what of this could we automate? Like I already have a plan built. Um, at the point where I'm like, great, now I have to spend the time to copy and paste it into our social scheduling tool.

Like that's the part that I'm putting off because I don't want to do it, but it has to get done. So, like, could an agent in here like do that for me? As long as and see, this is where someone like you is going to be a much better user of paperclip than in what in many cases someone like me, because you will have already thought out that plan. So you're not relying on AI to kind of wing it. Uh instead, you're saying this is how you're going to do this task.

Right. So even in this, you know, ridiculous uh silly example, you can see here that it has spit back and said, you know, good meeting uh notes should include a clear agenda and the key decisions made tied to the person responsible to target it, so nothing's substitute the cracks, action items captures discrete steps. So it did basically, we saw we've now seen it do its first task, which was the whole point. So okay, so I mean, what else can it do? Like, because I look at that and I think that that's a good example, but did I need to set up paperclip to get the answer to that question?

Because that feels like if I already have the co-CEO set up in Gemini and Claude, why wouldn't I just ask the skill in one of those systems? Why would I need to set it up in paperclip to do that? And again, granted, it's just an example, but I feel like it's helpful for us to sort of like wrap our heads around like why use this versus what we already have. And in a lot of ways, again, with things like dynamic workflows in Claude, you might not need this at all, because a dynamic workflow, as long as you've built the infrastructure around it, can do exactly what this has just done without all the extra overhead. And that's why this is such an interesting space in AI, because and we talked about this on the podcast.

Once you see a project like this spring up and the community run with it. If it's a good idea, the big vendors are like, well, we can implement something like that in our software. And then you, if you are, instead of being on the very leading edge, if you are just an early adopter, you can say, okay, I'm gonna wait and see if Anthropic or OpenAI or or Google implements something like this, which they have. Okay. So what would be a complex project where you would have you yourself, if you knew the uh shape of it, would have lots and lots of sub agents, lots of multiple steps, you would need lots of different pieces.

What kind of project would that be? Um I mean, I can see, I don't know if I'm like leading into what you're hoping I'm gonna say. Um, you know, one of the things that we talk about is live events for ourselves. And we all collectively have a working knowledge of what goes into planning a live event. We've done enough uh, you know, workshops and uh, you know, we've attended events as speakers that we understand the amount of work that goes into putting on a good event.

Um, you know, I'm not worried about like the workshop itself, because like we could do those in our sleep at this point because we do them so often, but it's the event itself, it's the experience for the user, and that is something that while we know what it entails, nobody on the team has time to execute against it. And to me, that is something that, you know, if we are looking at, you know, what are the best practices from the lens of um an event coordinator or an event director? Like, what are the things that we, as someone who doesn't do it all the time, don't think of that we should think of, or who on our team has time to do the research and actually talk to vendors and be like, hey, you know, we need this, this. This is our price. Like, I want what I would love in a perfect world is to say, here's the budget, here's the timeline, here's the amount of resources we can spend on this.

Someone go make a plan that fits all of that criteria and just hand it back to me so I can approve it. Like that's my perfect scenario. That is the kind of thing that at a small company you just can't get because everybody's doing everything. So you yourself like are participating in it, and it just you're you're in the weeds. And so that's the kind of like more complex project that for our company specifically would be something great to get assistance on.

So to do something like that, um, we might have something along the lines of a let's put up a new project. We'll call it work, you know, workshop planning. And planning a metro Boston area trust insights workshop. Uh, this is in the planned mode. We're gonna hit create the project or create the plan.

Uh we're gonna make sure that our first issue on this should be essentially that miniature project plan that you just talked about. Um, so this would be uh workshop planning, and we're gonna assign this to the CEO because the CEO then is going to make hires hires as it thinks through what is it that needs to happen. So we're gonna be planning a workshop in the Metro Boston area within 25 miles of downtown Boston, accessible by public transit for 25 to 50 people. We need to know where we should have the workshop. We need to know what venues are available.

We need to know what they cost and uh if the cost is inclusive or exclusive of things like catering. We need to know what kind of public transit is in the area. We need to know how far out we need to book. We need to know what kind of Wi-Fi is on site for a workshop because this is a workshop that will require people with laptops and high quality internet access. We need to know if the facilities have robust electrical power to make sure that attendees can plug in their laptops for doing AI work.

We need to know what kind of insurance a venue might require. We need to know what hotels and lodging are nearby if people are going to be staying. So that's all the just the requirements alone of what we want to know. Next, we have to give it instructions like any good thing, like uh any good prompt. As you resource and plan, consider hiring research agents that will go out and research all these things and project managers to consolidate and synthesize results.

When you get conflicting results, the majority result wins. So if you have five reports and three of the five have one fact, that is the going to be the threshold for confirmation. Your output is going to be a CSV file of the venues with all the listed issues as columns. It will also be a markdown document with no tables in the markdown that we can read and review. So already we've got a pretty decent issue here.

Now, anything else that you would throw in here, Katie? Um, you know, if given more time, I would give it a budget. I would say, and we can't, you know, what we could do, we could say, you know, our profit margin has to be at least, you know, you know, X percent or something like that. Let's do it even better than that. Let's have three different types of venues low budget, middle budget, and high budget, where high budget is going to have a cost of 250 to 500 per attendee.

Mid-range budget is between 100 and 249 dollars per attendee, and low budget would be 99 per attendee or less. You'll need to factor this into your research and do the computations so that we have different options and all the venues you research will have that budget information. If you can't find budget information for a venue, you must exclude it. Yeah. What else would you throw in?

Um, you talk through a lot of the details of the user experience itself, resource and plan. I mean, those are the basics because even looking at that, like it's just three paragraphs, but anyone who knows the amount of work that goes into finding all of that information, putting it together to make a singular decision, like that's weeks, if not months worth of work just to do that research, let alone call up a venue, say, Hey, you know, do I have to supply my own forks or can you supply forks for me? Like, you know, there's a lot of information. Um, the other thing perhaps we might add in is, you know, how well, I guess I don't, and again, this is like, I don't know. Do it, do we need to add in, you know, things about accessibility?

Um, you know, should we're talking about public transportation, but like how accessible to anyone should this venue be? Does it have you know any kind of other, you know, hearing or sight or you know, other impairment assistance on site, or do we have to provide those things depending on who's attending? Like, that's the kind of stuff again, we need to make sure that we are, you know, thinking through those details. Um, which is why, to me, if you're just winging it, you're doing it wrong. Mm-hmm.

Exactly. Uh Kimberly chimed in uh about the registration process, marketing and promotion, uh, website updates. Are there speakers at this event on Amazon? Yeah, Kimberly, don't get me started on how much work this is, which is why I've been putting it off, and the team is very upset with me. So now here's the next question.

Um this would be okay for a JIRA ticket in JIRA, right? This would this would do a pretty good job. This will not do okay in paperclip. And here's why. You do not get a chance to ask questions.

So one of the things that we teach in all of our workshops in AI is we say, well, the top prompts you should have is ask me questions. Do you have enough information to successfully complete the task? You cannot do that in here. Whew, that makes me nervous. Right.

So what we should do is flip over to a regular LLM first and use a prompt optimization framework of some kind to take the raw input and refine it and do the asking questions here first before it goes into paper clip. Because once you kick that off, it's going to kick off a bunch of agents. It's going to kick off a bunch of other things. It's going to use a lot of tokens. And my thing, and this is this is straight from Katie is it is better to spend more time planning and then you kick it off once, knowing that what you put in is is bulletproof, as opposed to going, oh crap, I forgot this.

And then you basically burned a hundred million tokens and you have nothing to show for it. Yeah, it's like I cannot stress that enough that all of the planning up front makes the execution that much more, you know, streamlined and seamless. It's the same thing with software development. And I've talked about the arguments I would get into with my software developers where they just wanted to push the buttons and do the thing and build the thing. I'm like, how do you know what the thing is?

Oh, I just know, I'll figure it out. I'm like, this is why we never hit our deadlines. And it's, you know, this feels like a really straightforward example. But as we're getting into it, and as you know, Camberly pointed out, there's so much more. We were just talking about the venue itself.

We haven't even thought about promoting the gosh darn thing, which is what I started with in the preview example, saying we have all these courses in our academy that we don't have time to promote. So, like, how do we have time to suddenly promote a workshop? It's a lot of things we need to think about. And, you know, I can't think of a single team that I know of right now that has people sitting around looking for more stuff to do. I said I personally can't think of one.

You can probably think of some. Oh, for sure. So what we did was we put this through a prompt optimizer. Um, and it what it does is it aligns it to the 5P framework by Trust Insights, which you can get at Trusted Sights.ai. But this is how we'll get better performance out of paperclip by making sure that it has some kind of framework.

So we can see here, what is the purpose doing this research? Who are the people, right? These are the who are the what is the process? Low there's our budget information, our triangulation phase. What's the platform, right?

Here's the outputs that are required. How do you know that you've done you know what's your definition of done? And so all of this now fits into a lovely thing, then this is what goes into paperclip. So it is not the um the winging it version. If you want to not burn 100 million tokens for no reason, you have to do the prompt optimization outside of it first.

That is such a big pro tip that if people don't take away anything else, take away that, please. Um, because to your point, and this is now I'm like, I have anxiety, is like once you start it, like it just goes until it's done. And so if you have not given it a good defined definition of done, it may go forever. Like, can you when when it starts, can you like check in and be like, hey, how's it going? Like sort of the same way you can with Claude.

You can be like, hey, how's it going? And it will show you all the steps, but like, are they gonna be meaningful? So as it works, you'll be able to see the activity and the related work and what it's doing along the way. Um so and and there is uh in the task itself, in the project itself, you can select the task and and hit the oh crap button and and hit pause on it in case you're like you're going off the rails, stop, stop, stop, time out. Um which is interesting, provided you're paying attention, because the whole idea behind these tools is that they work autonomously and you can go do something else.

But this to me says you have to be paying attention, like that's definitely conflicting in terms of what it is meant to do. When you think about JIRA, how much of Jira is hands off. I mean, based on how I've seen it used, none of it. It's all very manual and painful, right? So if so, if you are the PM, zero percent of Jira is hands-up because you got to be monitoring what's going on and making sure if you're the CEO, Jira is 100% hands off because you have a PM who's just doing the thing.

In this case, this uh paperclip system, the CEO there is the PM. So theoretically, but not in practicality, theoretically, it is hands off in practicality. If your your your CEO PM there is no good because you don't know what a PM does, this system is just gonna be an expensive disaster. Yeah. I did, you know, it's not to say there's different flavors of project management, but a better positioning for like when you're setting up paper clip and it's like, you know, build your CEO.

It should say build your project management manager. That would be a much better resource. Because this sorry, go ahead. It doesn't matter. It doesn't go into the whole tech pro, you know, agency that future fire everybody narrative if you say this is Jira.

Like you're not going to get 100 million dollars of investor funds if you say I just made Jira. If you say the most important person in this is the project manager, not the CEO. Right, exactly. Now, in this, you know, so as we're talking about designing these projects, like could you as the human build in checkpoints? So, like we just built in like go do all of this stuff.

Could we say after you finish this part, stop and let me check the work? Stop and you know, along the way. Like, is that something that you could build in? So, yes. Like Jira, you might break it down into 40 different issues.

You might say, You're just gonna research this, you're just gonna do this, you're just gonna do this, and and you break it down into individual issues, and you issue tickets out for each of those little things, and then you, the human can either consolidate them or tell the the consolidation agent, hey, review uh tickets seven through 14 and come up with a consolidated report from it after I, the human, have inspected the outcomes of ticket seven through 14. Again, just like Jira. I think depending on the nature of the project and how skilled you are with a system like this, that's probably a good way to start, just so that you can see, like, oh, this is the expected output, or this is where it goes sideways. So let me correct for that. It's like learning any new system.

You want to take those baby steps before throwing everything into it. Even if you have something that's really well prompted, you probably still want to build in those checkpoints because all of this costs money. All of this cause tokens usage at some point. You may not see it right away, but it at some point you'll be like, Oh, where did this like six thousand dollar bill come from? Mm-hmm.

Yep. Or the half a million dollar bill or whatever it was that was in the news this past week. Um and I gotta tell you, nobody wants that. People like surprises, you know, flowers, balloons, candy. That's great.

Six thousand dollar bills, half million dollar bills. Nope, not a great surprise. Nobody wants that. Um the other two things in here are goals and routines. Goals don't work yet.

Uh well, moving on. What those are though is either the high-level goal of a pro of a set of projects or from an architecture perspective, they're kind of like system instructions, right? So it was what the idea and then the routines we scheduled tasks. So if you had your virtual thing assembling and sending you out a status email every so often. Um so that's paper clip as it is right now.

Do you need to use this? Absolutely not. Um and here's why, because uh as we said at the top of the show, this is good to see, but there's nothing here that you can't do in clawed code or Hermes agent. Uh, and arguably in a system like Claude Code, you have more natural interfaces for you know, stopping and going, Oh, I should probably review this, right? The the that's just open code has a new desktop app out.

Uh OpenAI Codex just released a refresh. Uh Google's anti-gravity just released a refresh. And all of these give you the ability to do the same thing, but with a bit more uh work on your end of having that consolidated memory system. What does make this and things like Hermes agent valuable is that consolidated memory, but you can get that with add-ons to the existing systems without having to buy anything new or install anything new. Which is a great segue into our buddy Brian's question, which is what do you think it will take or how long will it take for these agentic systems to become predictable and safe enough to run in larger organizations other than tech orgs?

Um predictable and safe are kind of the most important words there. Right now, this system is highly unpredictable. Uh so it's got that's kind of a non-starter. Uh and the fact that I have to install this on a VPS means it's not safe yet either. I think what it's it's a it's a bunch of different things.

It's it's back to the 5P framework in a lot of ways, but more importantly, if the underlying agents don't have good governance, then the paperclip routing system, the JIRA on top of those agents isn't going to enforce that. And I was gonna say something very similar that the software is never the thing that you should assume is the safe or predictable thing. Software is volatile, it doesn't matter how long the software has been around. Like, you know, to pick on something like Marketo or you know, Salesforce, that can go sideways. It doesn't matter how long it's been around.

Um, if you don't have good governance, if you don't have good guardrails, if you don't have good requirements going into any project, regardless of the software. Think about like, you know, the JSON language or GitHub or repos, whatever. Those can go sideways if you give it the wrong instructions. And so that's not to say we shouldn't wait for these tools to put better guardrails in place. But at the same time, don't wait for these tools to put better guardrails in place.

If you have a use case that requires you to use one of these tools, you, the human, this is where we still need human in the loop and human oversight, you're the one who makes it predictable and safe. Don't wait for the tool to do it for you. Yep. And the other thing is on the predictable side, this is this goes back to level one stuff, which is in your prompt optimization framework, whichever one you're using, it has to be tuned for a genetic work. And that means, among other things, when you're optimizing a prompt, you have to have a very clear priority system.

Uh what is it, a week ago in our analytics for marketers Slack group, uh, one of the folks in the group shared a uh clawed prompt that was just not working. It was it just kept blowing up. And I took a look at it. And there were nine different top priorities. Like this is the most important thing.

This is the most important thing. Then you know PowerChaper, this is the most important thing. It's like the Incredibles, or if you remember Syndrome, when he goes, When everything is special, nothing is special. Well, you have 10, you know, nine top priorities, nothing is top priority. So when you put a prompt like that into a system like Paperclip or Hermes Agent, the model's like, so I don't know what the priority is here.

So I'm just gonna kind of wing it. If you have a good prompt optimization framework, it should at the minimum either ask you, hey, well, what's the actual top priority here? Or if you provide enough detail, it should be able to infer, okay, the top priority is factual correctness, right? So if you read our our workshop planning prompt, factual correctness is probably the top priority. Um, so no matter what else the agent does, it has to get real data to be able to help us plan a workshop.

You know, it's interesting that's something that I have so strongly built into my Claude co-work sessions that it's like, hey, so I didn't have this, so I took a guess, but you're gonna want to double check this because I don't know that it's correct. And I'm like, thank you for calling that out. Like, great, I will go check that. Not a problem. And if I I'm like, oh, here's the real information.

It's like, great, I'm gonna go back and I'm gonna fix everything. So hold on a second. Like, just give me a second, I'm gonna fix it. But now I have all the information. Like, that is such a big deal for any of these systems is the factual correctness because AI hallucinates with authority.

It does it with confidence. And if you don't know any better, you're gonna go, you know what? That does sound right. Um it might not be. You want to make sure, especially if you're making explainer.

Oh my god. You want to make sure that if you're especially making financial decisions, like in our example, you know, we're planning a you know, in person workshop, there's a lot of financial decisions and risk involved with doing that. That as the person who is in charge of those things, I want to make sure it's as accurate as possible. And that, you know, what we didn't add in there is that there is a cancellation clause that is clear and you know isn't going to cost us more than we put in. Yep.

So again, going back to the prompt optimization framework, you uh this is where stuff from like level one and level two still matters. You should probably do if you've never done event planning, you should do a deep research project saying what are all the things can go wrong throwing a workshop and have the different tools like Quinn and Gemini and stuff come up with the laundry list of here's everything's gonna go wrong, do that deep research, build it, consolidate, do emerge on it, and then that becomes part of the prompt optimization that you then then put that into paperclip. So even though paperclip is a very, very cool idea of a control plan on top of autonomous agents, it is still all predicated on that foundation of is the data you're giving it any good. And I think that that's a really good call out too, because we talk about you know, do I need to use AI for this? Yes, it uses tokens and usage and cost money, but that scenario planning, AI doesn't get fatigued the way that a person does or doesn't introduce the um, no, I think no, that's totally fine.

Like if you say what is everything that could possibly go wrong, you should be seeing things like a natural disaster, you should be seeing things like you know, a citywide power outage, you should be seeing things like you know, somebody gave you a bad, you know, Yelp review and it took down the whole workshop. Like you should be seeing all of those things, but let the AI come up with all that stuff and problem solve for it. Like, if this is all the stuff, what do we do about it within the budget? Then that way you don't have to worry about that kind of deep research for yourself. Let the AI do it.

And you, as the human can then use judgment to go, yes, that's something I should probably factor in. Like downtown Boston, we probably don't have to worry about, you know, perhaps like Godzilla coming through, but we might have to worry about, you know, uh a hurricane or, you know, really bad torrential downpours or you know, some kind of you know, strong windstorm that takes out the power for the entire city. What do we do then? Yep, exactly. And again, if you're a lot of what I see as really terrible use of these systems is insufficient planning uh up front.

You if you don't think this stuff through with your level one and level two systems with your GPTs and gems and just your straight up chat, then you're not going to get good results. So the results you get out of them will be either vague or will be kind of weird, and which is fine if you're just doing it for RD. But if you're doing it to for anything that's production, take the time to do to work your way back through the ladder so that when you hand off this stuff to your virtual company, which is what everyone's building this thing as, you've got all these materials and you know, the skills built with all the background data that it needs to have. It's funny because the most common question is will AI take my job. And we keep saying, not if you're using your brain, not if you are developing your critical thinking skills.

And we just spent an entire episode talking about how important it is to have those critical thinking skills before you touch any of this software. Exactly. So go go play with this in a sandbox somewhere if you are inclined, right? It is worth playing with to understand what the technology's capabilities are today and where it's probably going. Make sure that uh I would strongly recommend that you use an agent that has, if you have the budget to use a local AI agent, something like Gemma 4 or Quen 3.6, so you're not paying token costs because your first few tries are gonna go wildly awry.

Uh and if you're you know spending uh clawed opus tokens, you are going to be very sorry very quickly. Mm-hmm. Again, it's the whole reason why do the planning before you even bring anything to the you know to the table with the software. Exactly. So that's gonna do it for this week's show and for paperclip.

Good luck. If you uh do want to chat about your experiences, Popeye, you know, our slack group and stuff like that. Say hi to us on LinkedIn uh and things, and we will see you all on the next one. Thanks for watching today. Be sure to subscribe to our show wherever you're watching it.

For more resources and to learn more, check out the Trust Insights Podcast at Trust Insights.ai slash TI podcast at our weekly email newsletter at trustinsights.ai slash newsletter. Got questions about what you saw in today's episode? Join our free analytics for marketers Slack group at trustinsights.ai slash analytics for marketers. See you next time.