Latest / Notion in Practice / Notion AI Custom Agents: What You Actually Pay For + The 3 Layer Guide (Agents, Skills, Guides) E09

Transcript

Notion AI Custom Agents: What You Actually Pay For + The 3 Layer Guide (Agents, Skills, Guides) E09 · Notion in Practice

• Tim Jeffries: Hey everybody. Welcome to episode nine of the Notion in Practice podcast. Today we are having a look at custom agents and ⁓ incoming costs. Jerwin: Mm-hmm. Tim Jeffries: works. It seems like there's still a bunch of confusion about what you pay for and what you don't. And we'll even sort of dig in a little bit to think about how can you keep the costs down? How can you keep them manageable? And how do you think about the decisions you have to make now that we're gonna pay for these things? Interesting that we're in this, you know, we're at this point, Jo, and where it's been free for ages and ⁓ we've all had a chance experiment test it out. It may seem like a long, long way away when they were launched and they were free for all this time, but it's just around the corner now and a few people are a bit nervous. Jerwin: Yeah, is. Yeah, yeah, myself included. And I'm interested to see your Tim, because I know you've got way more custom agents and agent builds that you've been working on. yeah, I think I've been to a lot of Notion events in the past couple of weeks. And the number one question I get with custom agents is the pricing, because sometimes when it's free, ⁓ you forget There's a model that works behind this and some things may be too good to be true, but a lot of efficiency gains through ⁓ AI. now it's sort think everyone's thinking, yeah, like when May come around or when it's come out today, ⁓ maybe we'll drop this when the pricing happens. This the things that you should be thinking about because these are the things that we are thinking of because you and me, Tim, live in Notion every day. So where did you want to start with, Tim? Tim Jeffries: think it would be helpful just to, I guess, get on the same page about how it actually works, because there's a bit of confusion, I think, about ⁓ do you actually pay for. So it's pretty clear that the only thing you pay for are the custom agents. So personal agent, little nosy in the corner, or when you go kind of full screen and you're using AI and you're chatting with it it's going back and forth. Jerwin: Yeah. Mm-hmm. Tim Jeffries: that's personal agent and that's just included. The enterprise search, the meeting notes, all of those kind of AI features are all just included in your plan. And so the incoming costs and the usage based pricing is all around the custom agents. ⁓ And so there's a bunch of things you can do within the custom agents to them more expensive or less expensive, but that's really what we're going to pay for. And so we have certainly been thinking for clients and for ourselves at SmoothOps, how do we... Jerwin: Yeah. Mm-hmm. Tim Jeffries: structure our usage of Notion AI so that we use custom agents when it makes sense to, but not unnecessarily. ⁓ saw a promo the other day and you know, there was sort of demoing, chatting with a custom agent about something. And I just thought, I don't think I'm going to be doing that, right? I don't think I'm going to be spending time with agents paying for the usage when ⁓ I just use the personal agent. So thinking that through, I think is going to be key. Jerwin: Mm-hmm. Yeah. Which I completely agree, yeah. Yeah, yeah, yeah. Now talk me through that, Tim. If I'm someone that's, like I've already dipped my toes, I've got a couple of custom agents running, but I haven't really dipped my toes with like Nosy and I have a couple of other processes or workflows that I'm running. Now that custom agents pricing of the usage has come across, where should I start? Tim Jeffries: Yeah, well, we've been conceiving of Notion AI in sort of three layers. And some of those layers are things that Notion have sort of given to us and said, you know, this is how it works. And other things we've sort of invented while we're trying to make sense of it. Obviously, we're using AI inside of Notion, but we've used it in Claude and ChatGPT and a bunch of other places for a long time. And yeah, so. Jerwin: Mm-hmm. Tim Jeffries: So this is kind of our model for it. Layer one is agents and agents are made up of two parts. You've got nosy, the bottom right hand corner, your personal agent, you have your personal agent instructions and that is all day, every day, no limits conversation with AI. Jerwin: Mm-hmm. Tim Jeffries: Then you've got custom agents at that same sort of tier thinking about agents, but they are different instances of AI, different sets of instructions. Those are the ones you're going to pay for, but those are the ones that can also be automated. They can work while you sleep. They can work in the background. They can be triggered by things like calendar events or emails coming in or people putting an emoji on a Slack message or any of those kinds of things can run on a schedule. And so that's the agent level. Jerwin: Mm-hmm. Tim Jeffries: And what we worked out is you don't want to load those up with too many instructions. You don't want to make those like massive, yeah, complex beasts. Really you think about them as the orchestration layer. And then the next layer down ⁓ what is being called skills. We were talking about them as playbooks ⁓ the early days, but Noshan has called them skills and that seems to be what they're being called across the board, right? You have Claude skills as well. Yeah. Jerwin: Mm-hmm. Cheers. Yeah, yeah. Tim Jeffries: And I think about those as almost like the SOP for the agent. So the agent, you know, gets told, hey, you need to use this skill. And we can talk about sort of ways of making that happen that are easier or harder, but it calls the skill, picks it up, runs it, and then puts it down again and it's finished. And so we a lot of work into those at SmoothOps and with clients where Jerwin: Mm-hmm. Tim Jeffries: You really think through, okay, this is the process we want to run. The agent needs to do all these steps. Here's the context for it. Here's the guard rails around it. Here's how we make sure that the output is really good. You can have human in the loop steps and all of the kind of agentic workflow stuff. You build that into the skill. And then the final layer, and this has kind of really just been emerging for us over the last probably month is what we're calling guides. And that are... Jerwin: Mm-hmm. Tim Jeffries: like reference information that different skills and agents might call at different points. So, I mean, it's always the agent that's going to call on the guide, but the skill might reference it. So a really good example of this is we have a voice and style guide at SmoothOps. And anytime we're asking agent to just in the chat ⁓ by using a skill to write some content, either content is going to go public or an email or internal documentation, whatever. Jerwin: Have a good Mm-hmm. Tim Jeffries: We want to give it guidance about how it should sound, know, the way to phrase things, the style of language, whatever. And so there's a voice and style guide, one source of truth for this information that lives in the workspace that gets referenced a lot. So those are the three layers, agents, skills, guides. Jerwin: It's awesome. If I'm a client and I've got of three of these set up, say I've seen actually someone had ⁓ a custom agent set up and don't actually have a personalized instruction set up yet. like everything was built primarily straight away through ⁓ went the notion custom agent UI. They started building the constructions through there. But then they didn't have one. Tim Jeffries: I am. Jerwin: no Notion AR skills or no personalized instructions, do they take a step layer back or what do you suggest? Tim Jeffries: Yeah, I think that person's in for a big bill. And I mean, I guess it's like most tools, right? If you haven't stopped and thought about how you're using it, then ⁓ unnecessarily cost you. I think everybody needs to start with personal agent and so much so that, you know, like with their clients, I walk them through the process of building their instructions and give them tools to do that so that they have this experience of working with Notion AI that they've got this really clever assistant that does a whole bunch of other jobs. Only after we've got that working do we then think about skills. And then Jerwin: Yes. Tim Jeffries: After that, we start talking about custom agents because custom agents are really the automation layer, right? They're the bit that, yeah, okay, in the middle of the night, we want this thing to run so that when we show up for the meetings the next day, you know, there's an agenda sitting in everything, all the projects are up there. Exactly. Yeah, yeah. Jerwin: It's ready. Yeah. You don't have to wake up in the middle of the night to do that. Right. Yeah. Yeah. So could we, let's maybe we jump into it now. Like, um, um, do you, are you comfortable with sharing some of your gold Nice. Uh, listeners, um, listening, make sure, um, you check out the video version of this podcast because now we're going to be sharing screens. Tim Jeffries: ⁓ Yeah. Okay. Yeah. So this page, this is an internal page in the SmoothOps world around sort of like the index of our skills. And what we've done is, I don't know, I think we've got sort of somewhere between 30 and 40 at the moment, different skills. We break them up into these different areas of the business and ⁓ they get created, you know, for different purposes. And some of them we've talked about before, you know, I think this client catch up one was Jerwin: ⁓ it's fine on my end. Mm-hmm. Yeah. Very nice. Tim Jeffries: one of the playbooks that I demoed really early on when Notion AI first came out. ⁓ Yeah, and then other ones we've kept building as time's gone on. So pretty much now what happens is I'll be working with a client and running a project or working with one of the team and we'll go through a process and I'll realize, hang on a second, like that's something that we do all the time. I should get really super clear about what that process is, understand how I would do it manually and then Jerwin: One of the early episodes. Yeah. Mm-hmm. Mm-hmm. Tim Jeffries: run this build an AI skill skill to help write the skill that's going to do that in the future. And so this client documentation builder is exactly that. The other day I was making a video and making a one pager for a client about how something we built worked. And then I realized, hang on, I'll do that all the time. I'll have a big chat with the client about how it works in a video call. Grab the transcript from that. The AI can see the Jerwin: skillsception. Yeah, nice. Nice. Yeah. Tim Jeffries: actual tool itself that we built because it's sitting in Notion. I'll talk to the AI about it and then I'll run this skill and it will churn out documentation for the client that is like 90 % of the way there. It's incredibly good because it's got so much great context and the skill is very prescriptive about what we want the documentation to look like. Jerwin: Right. Mm-hmm. So then after that, you come in and you basically fact check the fact or team check everything that the skill builder creates, right? That's awesome. Tim Jeffries: Yep. Yeah, I mean, all the AI outputs really need, I think that they're still at a place where we need to be monitoring the output. yeah, ⁓ there's a governance layer to all of this that maybe is a whole nother episode all of its own, but that is about, okay, sure, you build a skill and it's the AI's responsibility to execute that. But who is checking that? Who is not just... Jerwin: Mm-hmm. Yeah. Mm-hmm. Tim Jeffries: like having a look at it, but who is holding it to a standard? Which person in the organization is responsible for that? It can't just be a general, know, ⁓ yeah, we all need to look after this. needs to be, it's someone's job to do this. And they need to be someone who can tell if the output is good enough or not. They need to hold it to a standard, just like you would with an employee. And who is checking over time that this skill is still the right skill and that what we need from it is what it's delivering, you know? How do we know that? Jerwin: Yes. Mm-hmm. Yeah, yeah, yeah. Tim Jeffries: The org and the world haven't changed a little bit and the way we work hasn't changed a bit, but the skill's still doing it the old way. So there's all those kinds of considerations. Jerwin: So what are you doing in that process right now? you going through, say for example, this client documentation, you're going through, you check everything, you check all the data sources are correct. And then do you use the, well, I'm assuming the verification feature before you share it with your team to use it for themselves or is there? Tim Jeffries: Yeah, I think Notion's verification feature is a pretty good one there. ⁓ These are issues that we're really only beginning to think about, but ⁓ yeah, I think all the skills really need an owner and they need a cadence, a verification cadence where you say, okay, like every 90 days or six months or whatever it is, someone should stop in and have a look at this thing to make sure that it's doing what we want it to do. And it might be pretty simple or yeah, might be longer term. Jerwin: Yeah. Tim Jeffries: But let's have a look here. I want to show this lead follow-up skill. Jerwin: I like this. Let's have a look. Tim Jeffries: Yeah, this is, apologies if you've gone through the SmoothOps lead flow process before, and it wasn't as good as it is now because this skill has made it a whole heap better. But basically, you we run a consulting agency, so we get leads. ⁓ People will reach out to us directly or through Notion or whatever. There's a bunch of ways that they might land on our plate. And they appear in a board ⁓ internally. Jerwin: Mm-hmm. Tim Jeffries: mostly because of this add new inquiry skill that gets triggered that puts them into our notion world. But when they get there, this is how we sort of centrally manage the initial follow up process. And so someone will go to the project that's been created with all the context from the email or the web inquiry or whatever it is and say, run this lead follow up initial response. And if we just have a quick look at what it does. ⁓ So it's got a scope and sort of describes what its job is. Jerwin: Mm-hmm. Mm-hmm. Tim Jeffries: We don't use this for people that we know or for referrals or like leads that are really warm that we've got some sort of connection with. A human just, you know, whoever writes, whoever knows that person writes the email themselves. But in this case, we don't know these people and we just usually got a little blurb. And so those are the sources. Then we say, okay, yeah, where is the AI getting all its input from? Jerwin: They're full cold late, essentially. Mm-hmm. Mm-hmm. Tim Jeffries: where's its context coming from. And so you can see here, either pulls it from the project page, from the person's page, it needs to know what their email is. It looks at our engagements where there'll be a summary, or there'll actually be the exact text of the email or the web form or whatever they filled in. And then it starts pulling guides, right? So it's pulling the voice and style guide that I was talking about before that says, is how we talk. ⁓ Jerwin: Mm-hmm. Tim Jeffries: but it's also pulling things like, and your marketing brain will love this, how we talk about smooth ops. So we have some internal language about like, this is how we see ourselves and this is the way we talk about what we do. Then there's a client intelligence section. So a whole thing about ⁓ who our ICP is, what pains they have, know, how does their... Jerwin: Yes, give it to me. Yeah. Mm-hmm. Yes. Tim Jeffries: buying journey go like what is the messaging at this point that's actually going to be helpful and will resonate with them rather than be kind of jarring and misleading. ⁓ And then our internal process about how our sales flow works, like, you know, what things we actually offer, what the packages are, all of that sort of stuff. And so the AI gets these four documents that really help it understand us and SmoothOps and what we do. ⁓ So it has everything about the client that we know. Jerwin: Mm-hmm. Mm-hmm. Tim Jeffries: everything about what SmoothOps is that we know, and then it can start thinking about what the process is. So that's a critical piece of the puzzle is give it enough context. There's a few edge cases. If there's no people relation, how can I write an email? All that sort of stuff. We define where the outputs want to go. So draft an email, never send it. Someone needs to look at this first. Jerwin: Yeah. Tim Jeffries: We update some things on the project to tell the project that we've sent this email. ⁓ And then there are the steps. And so we won't go through these in detail, but it's like literally step by step. This is exactly what you have to do. Old school SOP style. Yeah. Do this. Yeah. Exactly right. ⁓ And then there's a whole bunch of rules about the output. And you can see like, you know, this is the tone we want. This is the length, you know, all those kinds of things. If you come from this, Jerwin: Mm-hmm. After it takes all that context, will go through that step-by-step process. Cool. Mm-hmm. Mm-hmm. Tim Jeffries: source then do it this way if you come from that one. So a whole bunch of logic in there. Yeah, even like here's the ⁓ signature that's meant to go on the bottom of these things because they come from like an inquiries at SmoothOps. So all of that is built in. Jerwin: You speak with this sort of language. Yeah, yeah, yeah. signature. Mm-hmm. Tim Jeffries: Then there's an approval gate, so like some things to stop it. Like if the lead source is not this, then you have to stop and ask what's going on. So a whole bunch of safeguards. And then some examples ⁓ of different examples from different. Jerwin: Are these direct examples of what like you did in the past and you sort of fed it like these are some good emails that you've sent. Tim Jeffries: Yeah, think they've all been, yeah, they have, they've all been, these aren't real people for that purpose. But that's right, yeah. So I think it gives probably be like good and bad. So these are the examples it should follow. But then a bad example is actually really helpful for the AI too, like contrast, here's what it looks like to here's what it doesn't look like ⁓ and why it's bad. Jerwin: Yeah, no, I'll to cost that. Good examples of like email draft. Yeah. Yeah. Tim Jeffries: When I started including these in the skills, saw significant improvement in the outputs because you're defining like, you know, the bar and also like, you know, where you really shouldn't go. Yeah. It's super thorough. Yeah. But if you were going to employ someone in your business to do this, like you would have to think through Jerwin: Yeah, yeah, it's so thorough and so in depth at this level. Yeah, yeah. And so layered. Tim Jeffries: this level of detail to be able to train them to do the job. And that's, that's what an AI agent is, right? It is like an intelligence that's going to take on this role for you. like, it's amusing to me that people are like, I can't believe it takes so long to set it up. Like, well, if you want it to know what you need and you want the output to be good, then you have to do the work. Otherwise you're just going to get generic, you know, AI is going to make it up. Jerwin: Of course. you need to take that time. Yeah, I mean, but that's the problem with like general GPTs or general LLM models right now when how people use it that I've seen is, like, if they wanted to do this repeatedly, they'd either have one chat GPT or Claude chat, and they just trip through, can you send an email like, like we did before over time that chat floods up and then it loses a whole lot of context. And obviously it hallucinates. Or as here, like you do it all up front and, and, ⁓ Tim Jeffries: Mm-hmm. Mm-hmm. Yeah. ⁓ yeah. Yup. Jerwin: Does that sort of, did you, ⁓ does it learn itself over time sort of like how Claude skills work or is it, you prompting it to like within that chat functionality, I want you to remember this. So then therefore you either go into this skill and update it or you get your, ⁓ your AI agent to update that. Tim Jeffries: think that's an important consideration. So it's not self-learning. I mean, and generally AI has not. People sometimes assume that it will be because they're like, hey, this thing's smart. Maybe it should remember. And sometimes AI will, it'll tell you, ⁓ yeah, yeah, just, that's great. Thanks for that. I won't do that again. And my response in the chat is always like, yes, you will. I'm going to open another chat and you'll forget everything that's happened before and you'll do the same thing again. And it's like, ⁓ yeah, yeah, sorry. I know I do that. Yeah. Yeah. I mean, Jerwin: Terminator. Yeah, exactly. Tim Jeffries: One of things I've done is in my personal instructions, I forbid it from saying, ⁓ I'll do better next time ⁓ when we haven't gone and changed the instructions because it's, you know, because it's not true. ⁓ Yeah. So I think what's helpful for people about this is this is not a custom agent. This is a skill that one of our team will call at a certain point. And so we don't, we don't pay extra to run this, but we could. Jerwin: I'm agreeing with the other part. Yeah. Yeah. Mm-hmm. Yes. Tim Jeffries: have it as a custom agent, because what we could do is say, when an inquiry lands, run this follow-up skill. ⁓ Don't check it with anyone, just send it. And as a business, we might decide, you know, that's worth the cost. To be able to reply to someone as soon as their inquiry lands, is very likely to ⁓ help us get more clients, because that immediate response. Jerwin: Mm-hmm. Of course, yeah. Tim Jeffries: And yeah, there'll be a cost associated with it. And I don't know what it would be, but it wouldn't be nothing, right? It might be, you know, if we get a lot of inquiries, it could be 50, $100 in a month, maybe, because this intelligence has to gather all that context, understand everything that's going on, craft the email, send it out, log in the system that it's done that. And so I would expect, you know, that that's not going to be nothing. It's not a SaaS product. It's a bit of intelligence that's, you know, Jerwin: Yeah. Tim Jeffries: Really, it's replacing a person having to do that work. So anyway, this model allows you to have that as a skill that one of the team can trigger or choose to make it a custom agent. And so that's what our custom agents generally do is they don't have their own instructions. They just look out to a set of instructions elsewhere. So this is the instructions for my calendar sync agent. Pretty lightweight for a calendar sync. But what it does is it says, gives it Jerwin: Okay. Mm-hmm. Tim's calendar sync. Yeah. Tim Jeffries: tiny little bit of detail and then says follow the calendar sync skill. Jerwin: So you literally just add the skill within the custom agent. Tim Jeffries: And this is the skill, right? ⁓ And this is really important because everyone on my team that books calls with clients uses this skill. They all have their own custom agent because it needs permissions for their calendar. And so they have to set it up. But we all use this same internal set of logic because this is how we want, you know, our calendar agents to work. If we didn't do it in a skill like this, I'd have five, six versions of this, this set of instructions. Jerwin: Right. Very cool. Mm-hmm. Tim Jeffries: in people's individual workspaces where I couldn't see them. Uh-huh. It'd be a nightmare. Yep, yep, yep. And they'd all be working differently and one would be behaving one way and you're like, why is it doing that? Whereas now, if there's a problem, there's one spot to fix it. Jerwin: Because they'd all set it up a different way, right? Different sort of structure and schema and ⁓ that's clever. Yeah. Yeah. You fix one skill and everything, everything else gets fixed. And then I like the logic behind that because like you think about it from like, if it's like a calendar, sync skills, simple things like naming convention schema and within databases, like someone searching something up everyone, like I've seen some notion workspaces where just like everyone within an org is just naming shit randomly. Yeah. Tim Jeffries: Totally. Yeah. ⁓ yeah. Mm-hmm. Exactly right. Yeah. So maybe while we're here, we can just talk a little bit about the cost associated with doing these things. Jerwin: How much has it cost you? You're going to be a bit open-minded with that credit? $6,900. Tim Jeffries: Yeah, yeah. So let's have a look here. So this month so far, it's spent 7,000. Now the month is weird because ⁓ it's not like from the 1st to the 31st. It's this one has run from the 22nd of March and finishes on the 22nd of April. ⁓ And so to translate this usage into dollars, my understanding is that a thousand credits, these are credits, attend US dollars. Jerwin: Mm-hmm. his technicality. Correct. ⁓ Tim Jeffries: depending on where you are in the world. Yeah, so the US dollars. So it's like a $1.40 or $1.50 Australian, you know, for US dollar. that means that, you know, at this point in the month, I've already spent... Jerwin: Mm-hmm. Tim Jeffries: Well, I've had to buy like seven lots of credits to run this thing. And if they're, let's say they're $15 each. So I would think that this agent, like it looks like it might cost me 150 Australian dollars for the month. Jerwin: for Sony but. Tim Jeffries: And off it goes again, running in the background there. Yeah, I mean, it's not cheap. It's definitely not cheap. think part of the challenge with trying to work out like, it worth it? You know, the return on investment question, which everybody needs to ask. Well, the first thing I want to say is I think it's good. I think it's good that we have to think about this now because we have been in this kind of... Jerwin: What are your thoughts about that? Mm-hmm. think everyone's thinking about that right now. Tim Jeffries: play space where not just notion, not just notion agents, but like generally AI people haven't had to realize what the compute cost. ⁓ And it's, we get, we're feeling the pinch everywhere now, right? You ran out of Claude credits yesterday. Like I run out of them all the time and it says, you can't do anything until six o'clock. ⁓ Like they're real costs. It actually. Yeah. So I don't think it's bad. Jerwin: Yeah, yeah. To go outside and touch some grass now, Tim. Tim Jeffries: that we're faced with it, because it's not an unlimited resource. And so now the question becomes, is this worth it? And so for me, like this agent, if anyone books something or changes something or cancels something in my diary, then pretty much instantly my Notion Meetings database knows about it and sets things up. So it means that it manages the CRM. If someone books a call and someone new joins it, Jerwin: Mm-hmm. Tim Jeffries: It adds them to my people database. It connects them to the company. I have a record of that person that runs forever and I can see all the meetings that I've ever had with them. Super useful, right? Incredibly powerful for us as a business. I've got another agent that preps agendas. I'll keep this one. Yeah, yeah. But I can't think about it like a SAS tool. Yeah, yeah, that's right. Yeah. Jerwin: then therefore you'll keep this custom agent, right? Yeah, yeah, so you're thinking it from like a business value perspective. Tim Jeffries: I think the sticking point for a lot of people is we're comparing it to SaaS prices. So you think I wouldn't pay 150 bucks a month for something to like sync my calendar. And if it was just a SaaS tool, like an old school SaaS tool without AI built into it, that was, you know, essentially like an 8N workflow running in the backend, right? Like pro gaming do that. Then it wouldn't be worth it for sure. But if there's intelligence applied to it that makes Jerwin: Mmm. Yes, yeah, yeah. Tim Jeffries: you know, that makes the output essentially what a person would do. Well, then actually it's probably quite cheap, right? If a whole month of someone's work to keep this up to date, you know, however many hours that would be, it's only $150. That's pretty cheap. ⁓ Comparing it to my time, like, you know, if I think how much time and effort would this save me? How consistently does this work as opposed to what I do? What's the cost of me missing that and not doing it? Jerwin: Yes. Mm-hmm. Yeah, I like that way of thinking, right? Yeah. Yeah, yeah, yeah, yeah, yeah, yeah, you think about like. Tim Jeffries: This is totally worth it. But it's just different frames. So yeah, I mean, you can obviously do things to make it cheaper too. I could run it on a cheaper model and see how that goes. And maybe it would cost a whole lot less. I could break the workflow up into pieces. Jerwin: That will affect the custom agent pricing though, right? Selecting those LLMs with the custom agents. So if I choose Sonnet versus ⁓ Opus, then it affects obviously your credits within the usage. Tim Jeffries: Yeah. Mm-hmm. Yeah, that's right. And you can see also like experimenting with these smaller and open models that will be a whole lot cheaper. You can see they talk about like how smart are they? Yeah, yeah, yeah, yeah. So I think that's really helpful. But again, you need to sort of experiment and work out, OK, but is the model too dumb and it's going to make mistakes and so the output is not going to be there. You know, so we look at 4.7 which came out today. Jerwin: Mmm. Oh, right. I haven't seen this UI yet. Nice. Yeah. Tim Jeffries: It's super expensive, but it's super smart. Now, probably for my calendar sync, that's not really relevant. So I probably need to be experimenting with the Haiku model to see if that's smart enough to follow my instructions. So there's a whole bunch of nuance in here that most people haven't really played with yet. Jerwin: Super smart. Mmm. That'll be a fun episode, Tim. do maybe next episode is I'll go through our AI skills or custom agents and then you choose. then whether it's S tier, yes, use it for custom agent. Yes, you'll do it or no, don't use it. So for example, Tim, if I'm going to build a AI custom agent where I'm going to talk directly and use Opus, what would you rank that? want to talk directly on the chat. ⁓ Tim Jeffries: Run them on different models. Hmm Mm-hmm. Jerwin: And I'm just going to use Opus all day with it. Where would you rank that? Tim Jeffries: Well, because it's kind of, there's a wastefulness to it when we haven't been paying for it. Or even in, even with nosy, you know, if I'm running Opus 4.7 and I'm saying, can you have a look at this database and, you know, I don't know. Yeah. Pull every entry that's got like this string of text tucked away somewhere in, or, you know, some, some version that I can't do with a filter, let's say. You could do it with Haiku and it would cost. Jerwin: It's actually funny. Tim Jeffries: it would cost Notion a lot less, but it would cost the world a lot less, right? Like there's no need to be spending that compute. ⁓ But we don't, because it's hidden from us and we don't care that much. We should, but we don't. ⁓ There's a whole bunch of other things in here. I read a great article this week by Matthias Frank, and he talks about splitting your instructions for these models into like Jerwin: Yeah. That's AI mindfulness. Mmm. Mmm. Mmm. Yes. Tim Jeffries: the basically the parts that are needed that you can use different models for. So let's say in here there was something really smart that it needed to do and you needed to use Opus for it because the rest of the models aren't clue enough. That's fine, but just split it out and have it as its own piece of the puzzle rather than jamming it all in together. And a lot of our agents are built like that. You know, this would bring it down to the sort of the smallest of the skills down to its smallest parts. Jerwin: different use cases. Yeah. Tim Jeffries: and say, yeah, okay, this is just procedural. It doesn't need a smart model. We can use a dumb model and it'll go quick and it'll be cheap. And then, ⁓ hang on, we're writing a strategy doc. Okay, we want the best brain we can get here. Jerwin: Are you talking from the perspective of skills or from custom agents? Tim Jeffries: Well, he was talking about custom agents ⁓ in the article ⁓ because in the way he's thinking about it, the instructions are here. And that's kind of how Notion's built it here, right? The instructions are literally in the agent itself. But what I have found is that that skills model actually scales a whole lot better. It's a lot better collaboratively for teams. so, yeah, so in my world, Jerwin: Mmm. Mm-hmm. Yeah. Yeah. Yeah. Tim Jeffries: ⁓ That piece he was talking about is actually the skill bit and you want to split them out. Jerwin: Yeah, yeah, yeah, yeah. No, cause I was just wondering. Yeah. With the orchestration piece, it's like, obviously with the different skills or like the different custom agents, you just referencing like when you say split the, models, does it mean like splitting the custom agents functionality or was it within the instructions? If you're to be doing this use Haiku, which is like a lower, then if you're going to do some deep research. Okay. Right. Tim Jeffries: Yeah, so this is about orchestration. This is about, you know. Yeah, so you can't do it inside the same agent. You know, there's only one choice here. So it's going to use like Sonnet 4.6 the whole time. So if you want to use IQ for some and Opus for another bit, you'd have to split it out. And like an example in this case might be. Jerwin: So. whole time so that's it's called the custom agents. Yeah, got you. Tim Jeffries: orchestration of putting the entry in the meeting database and connecting the people to it. That could be done by Haiku. But if you want another step where a agenda gets prepared based on the context of the last meeting, then you'd use Opus. So you might do something like have the first agent put the meeting there, set it up, and then tick a checkbox or something on the database to say, Jerwin: Mm-hmm. Mm-hmm. Mm-hmm. This opus. Tim Jeffries: you know, please prepare the brief and then the other agent triggers and then sets up the brief. I haven't actually played with like linking them together like that, but I would imagine like that's a way to keep costs down, be more efficient about how your, yeah, how your custom agents are working. Jerwin: Yes. Fantastic. And ⁓ as you prepare for ⁓ custom agents and the credits system, is there a certain SOP, I'll use your terms, ⁓ that you're using from a higher strategic level to say, ⁓ so when May does come around or when you listen to this episode, viewers, ⁓ you can see a list of what you're using for, and then you have the, ⁓ Tim Jeffries: Mm. Mm. Jerwin: based on what I heard from you is the business value, and then who's using the custom agent, ⁓ and then the costs per month for forecasting perspective. So teams can allocate budgets for their custom agents. So for yourself, how are you doing that for your team? How are you thinking of doing that? Tim Jeffries: Yeah. ⁓ Yeah, so I did a bit of a cull. Obviously we've been playing with them, so we had all kinds of things in there. Now I've kind of pulled it down. This list here is actually, I think, pretty much all of the agents that I have running. And then you can see a few others from other people in our team. Jerwin: These are your odds. These are your top eight superstar custom agents. Tim Jeffries: Yeah, and so yeah, it's a question about like, where's the value? Like what's worth it? ⁓ Yeah, and then like the dollar figure, how much are they gonna cost? there's actually, if we have a quick look in here, there's a new dashboard for custom agents. This is a bug. This happens to me all the time. I don't quite understand what's going on here, but ⁓ like I looked at this three days ago and it was here. Jerwin: Mm-hmm. Right? Okay. Tim Jeffries: And then on that day, it goes, no, nothing. We'll come back tomorrow and it'll still be like tracking to this line and the, yeah, it's, I don't know what's going on. It's crazy. Yeah. But the numbers don't add up. I'm not sure. Anyway, um, yeah. So. I mean, I think probably for the first month or two, we were likely to sort of just keep experimenting. Jerwin: What have you been doing the last week, Right. You've doubled, doubled the usage there. 50,000 credits. Yeah. Right. Okay. Tim Jeffries: where the cost a little bit, you know, work out, ah, this thing's really not worth it. You know, we could just be a bit more disciplined as a team and do that manually. So we'll see, we'll keep a close eye on this board. You can track the individual agents themselves. Yeah, and then, yeah, it's a return on investment question. Jerwin: Yes. It is, it is a sort of like, well, this shows a nice view. ⁓ So could it's the notion, but I think what's missing well, like, yeah, there's runs completed there, but like what you said earlier, like, ⁓ I think this would become to the, the notion user itself is, ⁓ is from each agent, you sort of think deeply about the business ROI if there's no monetary return in investment metric there, like it's either saving you, for example, your, ⁓ your, ⁓ Calender sync, like what does that save you in time? If you were to do that yourself, you know, not going to get into your hourly rate, Tim, but like how much Tim's hours that saves. then you sort of, you know what I mean? It's a way of thinking about that. think that's a good way to think about it, but. Tim Jeffries: Yeah. ⁓ I think the other thing that has occurred to me through this process is that for a while now notions being that allows us to build our own tools. And it's meant that for lots of us who can't code and couldn't build an app, we can essentially build our own apps. That's what it's been for us. AI has come along and there's a whole bunch of other tools out there that are implementations of AI that don't allow you to build with AI. They just allow you to use kind of AI how it works in their tool. But Notion has a... Yeah. Jerwin: Mm-hmm. Yeah. how they constrained it, yeah. Tim Jeffries: But what I like about how Notion's done it is they're giving us the tools to build and use AI how we want to use it. So that whole skills guides, all those kind of frameworks, like we made them up because they fit with the rest of the digital operating system of our business and how we build for clients. ⁓ Someone else can do it however they want. So we've got the blocks. ⁓ And what's interesting about the custom agents is Jerwin: Mm-hmm. Tim Jeffries: Like this calendar sync, if we wanted to break this down, you don't need AI to do most of this. You could use Zapier, Make.com, N8n, all of those kinds of tools. Connect to the Google Calendar API, connect to the Notion API. You could do most of what this does without AI. But most people can't do that, right? It's too hard. It's too nerdy. It takes too long. But... Jerwin: Yeah, create some automation flows. Yeah, yeah. Mm-hmm. Yeah. Tim Jeffries: What Notion's giving people is not just agentic AI, but the ability to come in. I mean, you make a new one of these, and you literally tell the agent what you want. It builds it all in the background, connects it out to those services, and then uses a model to kind of run it. That is incredibly valuable. ⁓ And so we're not just paying for AI tokens here. There's a bunch more going on. Jerwin: Just, yeah. Mm-hmm. Now it's that overall contextual piece where you've got your life's work within your notion space and you can just add the database or add the pages. That's another thing I point out as well from observing your skills documentation. I saw a couple of workspaces the last couple of weeks is, and a few ⁓ other notion ambassadors or consultants have pointed this out, but. ⁓ your database management or the way you store your database is so crisp, Tim. I just like looking at the way you set up all your databases. ⁓ I think that's like the first layer of bright as well, like underneath. Wow. Look at that. ⁓ my goodness. is Tim Jeffries: Yeah. Well... It's even color coded at the moment because I've been messing around with Roe-based permissions in our team and working with contractors. ⁓ Jerwin: Very cool. Wait, what's green mean? That means you've validated it and an orange means you're still to check? ⁓ it's a permissions legend. Very cool. Tim Jeffries: There's a little permissions legend here, mate. Look here. Yeah, permissions is a whole nother episode. All of its own. But you're right. This structure, the clear thinking here is the foundation for everything else. Like you can't give, I thinking about it the other day. Jerwin: Yeah That's so cool. Yeah. Mm-hmm. Tim Jeffries: People used to ask, like before AI, people would say, ⁓ we really want to do automations. You know, can we set up these automations? And the challenge always was that most of the time, people hadn't thought clearly enough about the process that was being automated. They just said, ⁓ we really, you know, we want ⁓ new client projects to automatically appear in the pipeline. But they hadn't thought, okay, what are the channels that the new client projects could come from and which pieces of information need to be brought across? ⁓ And, you know, Jerwin: multi-infrastructure like, Tim Jeffries: Yeah, all that like real detailed SOP type stuff that hadn't thought about it. And so all that happens with automations and now with AI is just you go faster into chaos if you try and use those tools. Like they speed you up and, you know, automate the process of it happening. Jerwin: It's Tim Jeffries: it's ⁓ the cost you have to pay to be able use those things effectively. If you haven't done this bit of work, the rest of it is just going to be a nightmare for you. ⁓ And so it's a lot of what we do with clients is disappoint them when they say, Hey, you know, we want to build custom agents. And then we say, all right. Jerwin: Yeah. Tim Jeffries: but like where's all your data and how is it all structured? And they're like, ⁓ it's spread across the workspace in pages, but AI is amazing. It'll cope, you know? And of course it doesn't, right? They get really poor results because the poor AI is like scrambling. Yeah. Jerwin: Untitled pages across thousands of private pages. And then everyone has like their own individual private pages that no one else can see because it's private page. Tim Jeffries: Yeah. Yeah. Like five task databases for different functions in the workspace. And, you know, they say, can you make a task for this? And the AI puts it in the wrong one, but how's it meant to know, right? Jerwin: ⁓ man. ⁓ Yeah, doesn't mean like you've got 50 databases, so it'll just guess, right? Tim Jeffries: You've got to think clearly and then you've got to tell it. If you have a look in ⁓ my AI instructions. Jerwin: Al, hey Al, how's it going Al? You know I haven't brought out is my Al. Tim Jeffries: This piece here that I think we've covered before, but ⁓ yeah, is a really kind of critical component is this is helping the AI understand how the structure works. So we thought about the structure really carefully. We built it all out. And now Al knows where a project is. I don't have to tell him, et cetera, et cetera, et cetera. Jerwin: ⁓ right. OK, so now we're going to AI agents. And these are all the skills that are sorry, these aren't skills. These are the databases that the AI agent has access to for context. So if you ask it a HR question, it'll know which database to go to. Tim Jeffries: Yeah. Right. Yeah, exactly right. I now just think whenever we build anything in Notion, I'm building it for humans and I'm building it for AI. And sometimes how they need to access that is different. Right? Like what are the instructions for AI are very different from the instructions for humans. ⁓ Yeah. And so this page, it's not particularly pretty for a human to try and read, but it's been set up so that the AI will most effectively be able to like engage with it and understand it. Jerwin: Bye. Mm-hmm. Right. Yeah, exactly. Exactly. It's very, ⁓ Tim Jeffries: I just, while we're here, I wanted to point out this AI skills and guides. This is the piece of instructions that helps Al know what skill. Well, it's pointing at just our skills database, right? Just that one we looking at before that's got all the different skills in it. We were looking at the pretty human view. ⁓ But you can see the instructions quite simple. It's just like on every message that gets sent to you, Al. Jerwin: Very cool, okay. That's Al skills. Mm-mm-mm-mm. Right. Tim Jeffries: have a look at the skills database where Tim's a user, so where I'm like one of the users. So then the skills database can hold skills that I'm not trying to reference, that other people in the team have made and are not relevant to me. And then semantically match it. So there's triggers in there ⁓ and descriptions, and it's not looking for like the exact words, it's just trying to say the intent. So I can say, up on this client and it'll... Jerwin: Mm-hmm. Yes. Mm-hmm. Tim Jeffries: realize, ⁓ okay, the client catch up skill, that's probably the one he's talking about. It's quite good at that. ⁓ And so if there is a match, then just load it and run it. And that's 90 % of the time, it doesn't ask me, it just runs, because it sees it and knows. But if there's two and it's not clear, then it will ask me. And if it's still really not clear, but it's like he's asking me to do something, there's probably a skill for this, then I will match it against the groupings, which we're calling kits of the different. Jerwin: Yes. Very cool. I love that, and we're into all the guesswork. Well, not all of it, but. Tim Jeffries: Yeah, so it means that my kind of relationship with this AI is it's quite fluid, right? We just, I'm just chatting to it, asking it to do things. I know that there's skills, I can't remember the exact name and it just finds the right one and runs it. ⁓ And then same with guides. Jerwin: So question, yeah, with the guides, before we get to AI guides with Al's guides. So we've got Al here. I just wanted to drop this cool much. So shout out to the Korean Nation Ambassador, Issa. ⁓ What's your relationship like with Al? do you basically, Al is your PA? Because I've seen this as quite rigid with his access and his. Tim Jeffries: You love it. Jerwin: knowledge as you were scrolling through, ⁓ is L like the orchestrator business wide and underneath L if there's specific other cases, so like more sales or whatnot, you'll have a version under L or is L like more of an orchestrator across the whole business? Tim Jeffries: Yeah, it's a good question because it goes back to that sort of framework that we were thinking of before, like agents as the orchestration layer and then skills. So Al is my chief of staff. Right. Al does everything. There's no he is the whole C-suite, if you like. If I'm the director of the board, Al has all the other jobs. yeah, because all we need to do is say Jerwin: Yes. Mm-hmm. Cool. Hahaha. You Tim Jeffries: you know, we're working on a technical brief. And then he becomes a technician who gathers the context and then writes the technical brief for the developer to do the work that we want to do. you know, if like this is a skill that we wrote that reviews discovery calls to give the staff feedback about whether they've done a good job in a discovery call or not. So Al suddenly becomes like an HR expert. Yeah. So he picks up the skill. Jerwin: Right. Yeah. Mm-mm. Mm-mm. That's cool. Okay. Tim Jeffries: does the work and then puts the skill back down and just goes back to being L. Jerwin: Jack Jack of all trades now. ⁓ and you don't have to switch the, ⁓ personalized instructions a lot because the way you structured it is like. Al at the top and then everywhere. Tim Jeffries: Yeah. Yeah. I never touch them. Yeah. Yeah. Yeah, because this whole process, I mean, I haven't heard kind of notion really referring back to it in the way that they did in the original that you might have multiple personal agents. I haven't seen much of that talk for a while. I think people realize that that process are pretty clunky to try and pull off, like switching your instructions. You want one set of instructions that you use all the time and then, ⁓ you know, you might use custom agents. Jerwin: Yeah. Tim Jeffries: or you just might use this one agent. One thing to point out is that the skills idea has made its way into the workspace. I met with someone from the Notion team the other day who built this. Jerwin: within the notion AI. ⁓ wow, okay. So. Tim Jeffries: So personal agent, nosy, can pull skills. Now, you'll notice I'm not using it because we built a whole other slightly more complex way of doing it and managing it. But you can say this page is a skill. So client catch up like this. You can say that's a skill. And so then Notion AI will kind of know that it's a skill. And you get this little blue box at the top. Jerwin: And across all of Notion AI? Is it across Notion AI? Tim Jeffries: are all of the personal assistant, the custom agents. Yeah, that's not a thing. ⁓ But yeah, so when you are somewhere else in the workspace, you can say, and it appears as a skill up here like this. Jerwin: Okay. ⁓ so it's in there with any UI versus if you had like client catch up at page or pull up like thousands of pages. If you had that. Tim Jeffries: It's in the UI. Yeah, problem is of course that this doesn't scale very well. Like one of my clients has a hundred skills. There's 15 people, they're running an investment firm, they've got all these skills. You can't have a hundred of these things in here. So you need some dynamic matching, you know, which is what we were talking about before. Jerwin: The way you set it up. Yeah. It's, it's a lot more scalable with your approach, right? Tim Jeffries: Yeah. Yeah. But this is the first implementation of it. And I think for a bunch of people, this will be like a really nice way to get started. And then I'm sure Notion will continue to develop it. Jerwin: Yeah. Yeah. So you say a bunch of people that's like, if you're just starting a business, only have small team and I'm not just starting like you're scaling your business. ⁓ and you need to launch quickly that, that UI start there. And then over time, ⁓ build an owl or build your own version of our, like Tim that gets to this, this level 99 point maxed out, you know, Tim Jeffries: Yeah. man. We're just ⁓ so much at the moment. Every client meeting, you know, every time we take all this stuff into a new use case, we realize, ⁓ hang on, could extend it this way or there's lots to learn. ⁓ do you think we've covered custom agent pricing ⁓ a way that's helpful? What questions might people still have, do you think? Jerwin: Mm-hmm. Yeah, mean, yes. ⁓ I had another question is I know I've been getting quite a bit of it and I'll ask you is, ⁓ some people, sometimes they use notion AI and then, ⁓ obviously for within their workspace, but I also have how heavy power users that are using Claude and then then using that notion connector, ⁓ MCP connector, which also then grab create pages. we did. Tim Jeffries: MCP connector. Jerwin: briefly mentioned this in the last episode with, um, well, a couple of episodes back with Daria. Um, but, um, no, uh, what's the best practice you've seen using the Claude side with like Claude having access to Claude skills. And then, um, if I am using Claude and I have connected my MC pizza, my entire database or my entire notion space, um, what's the best way to reference these notion skills that you build or these custom agents. Tim Jeffries: Mm-hmm. Yeah. Jerwin: Well, it can't run a custom agent, but that knowledge, do you create a Claude skill within there, or do you just at mention those pages? Tim Jeffries: think if you abstract yourself like back a little bit, the skills in Claude and the skills in Notion, they are the same thing. They are a page of instructions for the AI to follow to get something done. So I was using the other day, connected to my Notion via MCP, and I was running my Notion skills from Claude because, and the reason I needed to do that was because Notion Jerwin: Yes. Mm-hmm. Yes. Tim Jeffries: at the moment can't spreadsheets, like can't read spreadsheets, but Claude can, does an amazing job. And so I was messing around with some data for a client and we were, but ⁓ bunch of the context that we needed was inside of Notion. And then a bunch was outside in these spreadsheets. And so Claude could span both worlds. And the that I wanted to follow were in a skill that we wrote in Notion. And so, Jerwin: Yes. Mm-hmm. Yes. So. Tim Jeffries: The MCP server just allowed us to kind of work between both worlds. Jerwin: Mm hmm. So you don't have to double your efforts, right? Tim Jeffries: and then Claude was able to. Yeah, I mean, it's just, think the best way to do it is to think about them ⁓ the overall digital operating system. It's not like Claude versus Notion. No, Claude is like really good for some tasks. And in those cases, you would ⁓ the MCP server for Claude and Notion to have this ⁓ relationship. Claude can do the thing it needs to do for you. But. Jerwin: Mm-hmm. Yes. Tim Jeffries: you'll run out of Claude credits and you won't be able to do anything until six o'clock. ⁓ Yeah, yeah, yeah. I mean, that's an area at the moment where the Notion personal agent is just amazing because ⁓ use it all day. Jerwin: Just the worst feeling out to be honest. Tim Jeffries: The other thing about the structure, the context that the AI gets, so I have been experimenting a lot for clients with how do you get best result of the AI? How do you get the most accurate, the most insightful? And MCP is pretty damn good, but it's nowhere near as good as Notion AI live. inside the workspace, traversing the relations, you know, way it sees and understands the workspace, especially if you've got good custom instructions, is miles ahead of anything else ⁓ MCP to talk to Notion. It doesn't do a bad job, but it's, a reasonable enough amount of guesswork going on that I wouldn't, I would never choose to use it unless I had to. Jerwin: Mm-hmm. Yeah. Yeah. I like how you said that before. And I have had similar experiences, like just testing out last week with Claude versus like having the notion AI skills or like the notion costs, not custom agents, but notion nosies would have better outputs because it would hallucinate less. it has a lot, I think it just has the best context because it's within the platform. then I tried it same way with the perplexity ⁓ connector. Tim Jeffries: Mm. Nighties, yeah. Jerwin: Widely different output and so I was a connector so I agree with you Key takeaways. Okay, Tim, what do you say? Tim Jeffries: Well, I talk the whole time, so you have to come up with the takeaways. Jerwin: Yeah, look, the key takeaway for me is quite simple. you think through custom agents, think through the layers that ⁓ Tim showed you today. So it really starts from the bottom up. If we look at our beautiful LEGO piece notion ⁓ in practice image, ⁓ it really in really well. That bottom data piece is your databases. ⁓ make sure that's all organized. Like he's so Tim's like awesome, ⁓ set up and it's color coded by, ⁓ we can go through that next. The permissions. didn't even know. Let me think about you could do that and then say who can access what, because obviously, motion, you're listening to this, please allow me to, to password protect my finance pages, please. I'll a lot more in there, but for now, nothing. ⁓ but yeah, that's the first layout. ⁓ that's my first takeaway. And what's that second layer, Tim, once you get your databases set up. Tim Jeffries: Yeah, then think about personal agent. Then think about that guy. Yeah, and the instructions. anyone's listening and they're struggling with their instructions, reach out to me. I've got a special present for you. But piece, ⁓ it everything. You know, the amount of people that I've been working with and I've said, make sure you set up your instructions and they don't. And then they come back in the next call and they're like, it's rubbish. We go through the process of setting them up and it takes... Jerwin: personalized agent, so this guy. Yes. Tim Jeffries: 15 minutes, 20 minutes, and then from then on, 10 times better. So that's the missing piece for a lot of people. Jerwin: Okay. And then on top of that, that neck, takeaway number three, and that's a very kind offer guys. Make sure you take that offer because I've played around with it a little bit and I ended up spending like four to four, maybe six to seven hours on my Sunday playing around with that. I'm not, I'm not even kidding. Oh, like, I know I need to drop six, seven, but actually spent like at least six hours on it. Like refining all my, um, personalized instructions. But now I'm like from talking to you, Tim, I'm like, because I have like eight, I have seven hours, but maybe I need one hour, you know? ⁓ Yeah, I do prepare per functionality. So like I have one for paid, but one for productivity, one for whatnot, because they're like different. But the next up is those skills, like AI skills. Tim Jeffries: Really, you've got multiple personal agents. Yeah, right. Jerwin: So. ⁓ Tim Jeffries: Yeah, yeah, once that once you got the personal agent working for you, then think about and I've met with someone just before this call today, you know, sort of like, how do you get started? I was like, just pick one workflow that you have that you do and then try talk to agent about creating a skill. And it's I mean, it does a much better job than I do at writing them. Right. So I get it to write them. Jerwin: Yes. Yes. Tim Jeffries: And I've got that ridiculous meta prompt that is like, build a skill skill. ⁓ But yeah, that's the way to do it. Because then you make it and then you just experiment with it. You say, OK, know, spin up a new chat and say, OK, run this skill. And then you see the output and then you work, ⁓ hang on, needs a little tweak. And then you tweak it and it just gets better and better over time. Jerwin: Yeah. Yeah, yeah, cool, cool. And then last but not least, what's that last layer? That's not the last layer, but it's the one that everyone's thinking about, What's that last key takeaway? Yeah, yeah. Tim Jeffries: ⁓ the custom agents, you mean? Yeah. Well, don't forget the guides. I think the guides ⁓ are one of those elements that really helps the output. Otherwise, what you end up with is skills that are huge because they've got all this reference information in them. And you've got sort of 10 different versions of that because it's in every skill. So yeah, any reference just put in a page and link to it. But yeah, of course, custom agents, when you decide this agentic workflow, Jerwin: ⁓ right. Yes. Sorry. Sorry. Tim Jeffries: needs to be automated, bring their custom agents in, you know, and obviously tweak the knobs to try and the bank balance. that's, I mean, our business is loving using them and they're absolutely incredible. definitely worth getting, definitely worth doing all the work to get to that spot. Jerwin: Mm-hmm. Perfect, Yeah. Yeah. Yeah. I like that, that approach. It's like a nice, ⁓ Lego, ⁓ build eventually get there. You'll have a nice, nice smooth operating system. Tim Jeffries: it. Good mate. Thank you. Jerwin: All right. Well, guys, make sure you message Tim for that little bit of a gift. ⁓ if you enjoyed this episode, once again, hope you ⁓ crazy with your OWL your custom agents or your personalized instructions or creating your guides. If you found this episode helpful, please leave us a rating on Spotify, Apple. I've got ⁓ quite a of you on Apple. I did not know that. Or leave a on YouTube. Do you have any questions? Either just reach out to us on LinkedIn, Instagram, whatever channel that you want, ⁓ A link is in the podcast description. Well, hope you guys have a fantastic day and we'll see you in the next episode.