AI in Context, Ep 02: Managing agents, for people who already manage people

Machine transcript of the replay, unedited. Times link to the moment on YouTube.

0:00All right, so I'm gonna get started. I'm gonna change this one up. Like I want to make this more approachable. I'm gonna ask you guys questions in the chat, like just one question in the chat. But other than that, it's gonna be about 15, In 20 minutes, today's goal is very much like yesterday, like last week we started on really understanding agents, what it means to get started with them, what it means like delegating tasks to AI agents, what does that look like? I'm gonna come back on those things, but today I really wanna give you more insight with regards to how can I, like I know how to deal with humans. Everybody here is an adult. You might be a young adult, you might be an older adult, but you're still an adult who has learned to deal with humans. And those same skills apply to agents.

0:52And so my objective today is to kind of give you the framework of how do I start delegating tasks to my AI agents? Because we're all going to become AI managers over the next three years. Like every single one of us is going to have this be part of our jobs. And today's my goal, the outcome of today is I would like you to be able to take a task that you have on your mind and delegate it to an AI agent. That is success for me. So the things you'll have ready, if you want to follow along, you don't need to. It's Claude or chat GPT. As I said last week, I'm going to be using Claude this week. I'm going to be going between Claude and chat GPT just because it doesn't really matter what tool you use.

1:40Have in mind one job that you do every single week and then if you have any notes from last week let me know and have those on hand. So like notes from last time you did it would be in this context I'm specifically referring to my example this week will be I want to go through all of my meeting notes like all of the transcripts of all the meetings I've had last week and extract the things and the engagements I took. That's going to be like the premise of today's like today's thing. I want to teach you how to do that, right? So in summary, there are four steps involved with this of delegating tasks to agents. Those four steps are A, I want to give the job to the agent.

2:30Two, I want to check what was done. I want to fix anything that went wrong. And then I just want to let it run, because once I have delegated the task, I want to set it up as a routine so it does it automatically next week. Again, it's this thing of compound. The more things I set up over time, the better my business runs, the better my life runs, without me needing to always be there managing everything. It's kind of the same four steps and like what's really funny is HR managers are very good at this. It's the same steps as delegating to humans, these things. So let me give you guys an example first, right? I wanted to tell my agent, this is the same example I used last week, I'm reusing it because I find it's a really good one, but I wanted more people to hear about my business, right? So I could have told my agent, be more active on social media. What an awful job to give an agent that has no idea where to begin.

3:38Instead, I gave it a very clear objective. I said, find the podcasts, email the hosts, and follow up with them, because I have the clear, actionable thing that I can follow along with, right? It's something I can measure, and I and see or know whether or not I achieved it. Now... To do this yourself if you want to follow along this is kind of what it looks like it's and as I said We were talking about like meeting follow -ups is I'm gonna create a new project in Claude and I'm gonna give the I'm gonna call that project meeting follow -ups and what I'm trying to achieve here is after each team meeting a list of Who promised what to whom by what that's really the objective here So, I'm going to go over and I'm going to take this meeting follow -ups right here.

4:29I'm going to go over to my project over here and I'm going to call it meeting follow -ups. I'm going to go ahead and take my text here and I'm going to put that in as a description. I'm going to remove the weird artifacts that agents do when they build these things. And there you go. I have my meeting follow -ups. I have after each team meeting a list of who promised to do what and by when. Now, what's really cool about this is if you have like a meeting transcriber, you can already at this point connect it to this project and have it automatically add all of the transcripts to this project. Now I'm not going to, I'm not going to show that right now because there's so many tools you have granola you have teams you have all of these fireflies etc all of these different connectors. You can send me a message if you want to like me to do like a short video on your specific tool but for today what I decided to do is I went ahead and created some mock like meeting notes for like weekly team meeting, another weekly team meeting here, another weekly team meeting here. More or less, this is basic stuff, but it's really about more the principle behind it and less about the specifics of what we're doing here.

5:46So are there any questions at this point? This is very much about structuring my work. I'm trying to bring it back to the basics at this point and say, okay, I've got this like blank slate of a cloth. I created a project for a specific task and now I'm showing you exactly what I'm gonna be doing. So like all of my routines, my schedules, my automations with regards to meeting follow -ups are going to live within this project instead of being spread across so many platforms. Any questions? Cool. Wonderful. So now that I've done this and I've created my project, you know, it takes two, three minutes to generate this and create it, well, then comes the question of checking, right? This is something I've learned through trial and whenever, it's if I ask my agent this like a very simple, hey, what did we do last week? More often than not, the agent is going to hallucinate something, or it's going to create what's even worse than a hallucination, partial truths.

6:52It's going to check the last time it got updated, and it's going to give me this approximation of what actually happened. Now, our goal is to avoid that happening. It needs to actually go and get the real information. And the only way you can do this is if you give it the proper context, right? And you make sure to check that you're asking the right thing. In essence, the best way to get better at managing agents is to assume that every single error is within your control, right? Like if you start saying, oh, it's because the model's weird, because the model is hallucinated, you don't have any more agency with regards to solving your problem. Instead, if I say, OK, I just need to work on my prompting, I just need to work on exactly what I'm asking in such a way that it's very clear when it is right and when it is wrong, I, A, have more ownership over the thing, B, I can be much more effective, and C, when I start letting it run by itself, I have like things that I can actually use for business and you know trust so there's a big element of that This is a story like I did I often ask my agent. Hey What did we do with X on this like in the past months and oftentimes it'll come up and I'll say like oh you Promise this in June you never built it that she's had ten weeks of silence and it's I feel awful because like like I feel like I'm being berated by agent for not following up on something, but I actually, I did my job, right? I sent her something in August. Might've not have been in a channel that my agent had awareness of, but it's like, that's what I mean in terms of making sure to set things up properly.

8:44So let's go back here. So how, in real, how does this look like? Well, I can go ahead and say to my agent, make the list from last week's meeting, for example, right? And I can go here, I can create a new job, and I can just say, hey, you've got all the context right now in your history, go ahead and make the list from last week's meeting. And at this point, it's going to go, it's going to figure out the date of last week's meeting, it's going to go through its context, it's going to understand, okay, what are we talking about here? and then it's going to create the list from last week's meeting. Now, this, right now, we're in the testing phase. We're figuring stuff out. We're trying to figure out what we need to fine -tune to make this really useful for me.

9:34So it went ahead, found the context from October 9th's meeting, and it said, here's who promised what, da -da -da -da -da. Who promised what? Who, Luke, Priya, Dana, Priya, Sam, et cetera. It extracted all of the things. And this is where I will use my own judgment and say, OK, was this actually what happened? Because I need to trust my memory as well of, OK, I remember the meeting. This is not exactly what was said. And this is my opportunity to start fine tuning, right? Because my prompt is very, very, my prompt right now is make the list of last week's meeting. It is a very, very simple prompt. But it is a good baseline to start saying, OK, okay, where do we go from here? So what's interesting about this is when you start giving it context, the agents have gotten good enough that they start really even raising things that you guys might have not mentioned or like were relevant in other weeks, but maybe not mentioned. These are really cool things that your kid starts saying, okay, okay, I need to fine tune what I'm saying and adjust, right?

10:41So, at this point, I've given it the job of checking through the things, I've checked to see if it works, and this would be my opportunity at this point to give feedback on this reply, in essence. And like, first thing I see is, this is too verbose. Give me bullet point lists per person and figure out next steps. I want this to be actionable. This is the, like when you're generating your routines and your schedules, it's the right time to actually work on the wording. work on making this practical and useful and so you don't rip out your hair when you receive this in your inbox every single week.

11:33Like, my rule of thumb is try to get it as short as possible. Like, if I'm gonna get an email from an agent, most people are very good at ignoring what agents send them because they're agents, because you're not gonna hurt an AI agent's feelings. And this is a deep empathetic element of humans. It's easy to ignore agents. And so what you need to do for yourself is make sure to reduce as much as possible. I want it as short and as brief and as useful and actionable as possible. Why? Because, well, if I do this properly for my team, for example, well, it becomes very easy to then have the agent send the tasks to each individual. And so if Priya receives just like this kind of summary after the meeting, this is so much more enjoyable than some AI slot document that is three pages long and goes into like deep articulated details about something that is never going to get read, right? Your goal is for it to actually get read by humans and so you need to respect their attention span and their attention span is about 20 seconds and that's normal because we're all exceedingly busy, even more so with agents.

12:52Any questions before I go on? Or any comments? Because you guys all were nodding your head and like, feeling the pain. Good, good, good. Nope. Awesome. So now that we've done this, right, we've checked it, we've gone and done that. One of the things that I want to be sure about, because now we're doing like very simplified forms is how do I know it's saying the truth, which is the right opportunity while we're building this prompt out. So just like as a reminder, because we're covering a lot of ground, the goal here is to have a system in place that after our weekly meetings, everybody gets their tasks directly given to them without any human intervention. Our goal is to have an agent whose responsibility is to take the transcripts of those meetings and delegate those tasks across the team right now in order to make this this useful for that kind of context I need to specify where my notes does it say that The objective here is to have an eval is to have an eval is an evaluation in order to make sure that what we're going to be sending out to our team is factual and based on the actual context and data we have.

14:13Figure out how to do this, knowing that we're going to send each person just their snippets, just their portion of the bullet points and their tasks and their engagements. So again, we're working on this like evaluation, I have my objective, I've created, I know what outcome I want, now I'm figuring out how do I measure whether or not we achieve this outcome. And we know that the frequency for like a weekly meeting will be after the weekly meeting, so I'll be once a week, right? While it's doing that, I'm just going to go down here. So this is like, I'm fixing it, right? I send it back to do this process. When people think about routine making and creating these agents, they think about this very dev -centric process that's like, you need to be a developer, you need to be a coder.

15:04Not at all. It's all about being a very effective and clear communicator. You need to get good at communicating to agents and AI what is your actual intention and how the AI can actually say that it has achieved what you have intended for it. So, for example, the slides or the Excaladraw you're seeing right now, initially I sent it to create this thing, and it created this super verbose 49 slide deck thing, which had context that was very relevant to me and my company, but no one would actually care or want to know about any of these things. And so what you have to do is you have to be very, very hard on the AI to really be like, hey, get to the essential, you know, cut the crap.

15:56Don't give me all of this verbosity and this like eloquence. All I need is what is this like, this kernel of truth at the center of it? Because I don't have the time, the energy or the mental space to take into account all of this context like you an agent does. Like an agent has no problem taking into account all of this context. Their brains are huge. Ours are not. Our focus and our attention is this, but we're much better at making decisions than they are. So let's utilize that. Let's be strong at that. So if I'm going back here, you see that this is a more complex command. Asking an agent to prove that it, like prove that it knows what it's doing is much more complicated than having it try to create something for you. And therefore, it's the most important part of the process.

16:52Rule of thumb, if an agent takes a long time to answer you, it means that you've asked a good question. If it's this kind of context. If it's taking a lot of time and you told it, go build my million dollar app, then that's on you. It has no idea where to start, so it's kind of doing anything. But if you're in this context of like, we have a very specific scope and we know exactly what we're doing here and it takes a little bit of time to do it, that means you've asked a really good question. And so it's like, it's a good, like, you should feel good when the agent takes two, three minutes, even though like, in your every day, having the agents take two, three minutes to do something might be very frustrating for your workflow. It's important to take it as the good thing that it is. So right here, I'm gonna close this like JSON that it just opened, which is just like a code for a very simple code format. and it's saying, okay, here's what happened.

17:46Several things in my last list weren't in your notes. So it made stuff up. I'll mark exactly which ones, then build a check that every line going to your team must pass before it's sent. So it's doing its own job right now. It's making sure that we have an evaluation framework that when we start running this by itself, it has checks and balances to make sure that what it's sending is correct. One mistake got missed, missed was allowed because the check was reading da -da -da -da -da. I'm not going to go into all the details, I'll be able to share this with you guys, but it's actually going ahead and saying it went through the evaluation and then it created a document for each person that only has, you know, call Mrs. Henderson to move the install date, the date was October 12th, there's no update in the notes since, reply done. Like, this would be a task that if I'm Priya and I receive this following the meeting, I can just say, hey, this is done, or I can do it on our task board or on our JIRA or on our Notion or whatever tool you use for task management.

18:51And like, we've been effective as a team then. Same thing, confirm new Henderson, da, da, da, da, da, da, da, da, da, da, da, da, da, and each person is now has their system. Each person has their responsibilities. Yes. Okay. And then it has it created some code here to check to see if what we're doing is is good because you don't want to overly rely on models to do this you also want to have what we call deterministic checks to make sure that we're doing things properly. Wonderful. Then, so this was very much up. This is what we just did we were reviewing it we were making sure have it go through and create the evaluation framework. I find my wording here much better. It's like the effect, you need to give it the objective.

19:41You need to figure it out. Hey, make sure that this is true. And then finally, it's going to be a question of then activating this as a routine. So right now, what I'm going to say is I'm going to go ahead down here. Apparently I'm approaching my weekly limit on this account and I'm going to say, let's make this into a scheduled routine that runs every Friday at 4 p .m. Our team meeting is at 2 p .m. Make sure to notify me if you didn't receive the meeting notes. So this is how I go from a chat to a agent because I'm taking my chat, I'm taking everything we've built together within this context, and I'm saying, okay, it is time to make this into a scheduled routine that every single week is going to be running.

20:40So at this point, it's going ahead, it's got its eval, it's got the claims, it's setting up the reminder, and it's just doing everything it needs so that within our meeting follow -ups, we're then going to have these scheduled tasks that are going to appear and that are going to be able to like summarize our email our meeting notes at this point. So there you go, Friday meeting follow -ups every Friday at 3 .49 p .m. Not sure why it shows to do 3 .49 p .m., but whatever. Then there you go, under schedules right now, I have the Friday meeting follow -ups. Under Friday feeding follow -ups, I have the instructions, weekly follow -up run for Northbank renovations, team meeting is every Friday at 2 p .m. Here's the steps, step zero, step one, step two, step three, step four, five, six. What's important here and the reason I'm showing this is how do I get it to have the actual data? Now, this is where here I would use my connectors. If I'm using granola, for example, I would actually say, hey, do you have granola under your connectors? Are you able to go get that data?

21:52This is where you check it. This is where you can see if my routine is working well and this applies a lot to like IT departments because a lot of people who get good at AI start kind of collaborating and helping it with a lot of this stuff because like it's job is not AI and from my experience, most IT departments don't want to deal with AI. because it's not their job. Their job is to make sure that our systems are operating so that the business can do its job with its tech, with its computers, with its accesses, with its yada, yada, yada. So AI oftentimes has fallen on like the shoulders of IT and IT departments, almost every single one of them that I've discussed with have pushed away that responsibility because they don't consider it their job. AI is owned most of the time by the executive function. So the leadership teams and heads of departments more than it is by IT.

22:54Why I'm saying this is because all of us, people listening to this people here are all like AI enthusiasts or people who care and feel the technology and know that it's important. Well, it's our job to understand how these systems and to be able to bug fix them. You can always ask the model to help you bug fix, but understand that this is where that information lives. Chat2pt is the same. I've got my notifications over here. I've got my connectors, so I know that it has all the information, and I have my general purpose settings of in which project is this, because I could change it to another project, et cetera, et cetera, et cetera. Last thing is, this is always the opportunity to choose the model. Right now, I'm going to use Opus for this. This would be a very good situation where Opus is the right choice because I don't want a stupid model delegating tasks to humans. I respect humans intelligence and therefore I'm going to try to take the most, like the most capable model that's within my budget to do the job.

23:59And I'd rather use something like Opus because it's running once a week. It's going through a meeting file. It's extracting the things like this is not a heavy task. This is not something that's happening every minute. And therefore, in that context, I'm not worried about this being too expensive. There you go, finally like this. And we're at the point where I'm pretty happy. I've scheduled this. I've let it run. It's a weekly meeting. And there's not much else to add to this. So I'm going to stop there. I didn't want to go overly complicated on this one because I think the principle is clear. Then it's a question of application. So I would like you guys to ask me questions and say like, hey, I've got this use case. How does delegating that task look like?

25:06Go ahead, Hal. I would be like something like the result there and it changes every morning what my competitors are looking at or maybe specifically what skill sets they're applying for. What your competitors are looking at. So you want a crawler that is looking at what do you mean by your competitors are looking at? Okay, you see it's important. Let's see Deloitte's idea. Some other big companies.

25:53What are their leaders hosting on them? What is important to them? Words. Are they talking about, I don't know, there'll be more data. I just want like a list of things we've talked about. Wonderful, I love this. Okay, so let's do competitive intelligence is what you're looking for. Yeah, thank you. Wonderful, so when you want to do competitive intelligence, let's take your example. So you want to figure out, your initial question was, let's put this to medium. I want to know what the competitors are talking, thinking about on socials.

26:48So because I can't I can't get in their head. I can't I can't know things that are behind closed doors. So I'm not doing espionage here, but I am doing competitive intelligence. So I'm figuring out what's what's happening. So there's a few steps here and I'll show you exactly how I'd approach this. So the first thing I would list is list out the people I want to track, right? And the way I would do this is I would go to something like this. Let's actually, let's go ahead and create a new project and call AI in context examples and random examples. Go ahead, and I'm going to say let me, so this is like, this is really cool because this is something I would do in real life.

27:44So let's go to here, connectors. Do I have access to Apify in this context? Add connectors, browse connectors, Apify, Apify, desktop, Apify, there you go, there you go. So why am I doing this? Well, you just asked me to crawl LinkedIn. Now crawling LinkedIn is an iffy topic, but it is technically feasible. So this is how I would do this. I would go ahead and take a crawler like Apafy. There are other crawlers that exist, but what I'm gonna ask it here is I'm gonna say, use the Apafy crawler and Apollo crawler crawler to find 100 people from the big four consulting firms that are active on either LinkedIn, Twitter, or any social we can crawl.

28:47So this would be the first step. I'm going to go and say list out the people I want to track, right? Once I have those people, what I need to do is create a script agent to look at all new posts one times per week, for example, right? Then I want to summarize what was discussed and put it into categories. Then I want to make a digest to send me every week. So at this point, I've actually gone through every single step of the process.

29:39Now, a lot of these things, so if we go back over here, what roles should they be? See, they're going to ask me questions. I'm going to go partners, directors, wonderful. Any practice, Canada plus US, yeah, that looks good. Post in the last 30 days. Yep, no problem. Yes, reveal labels. No, I just need the, I just need this. So like, I have the Apollo connector, I have the Apify connector. I do recommend, so in these kinds of like, I need to go find people, tasks, use the services that are specialized at that. Apify, Apollo, Clay, these are all tools that if you're in a big company, already you probably have access to, if you're in a smaller company, it's absolutely worth it to purchase these tools and use them. I think Apollo's like $60 a month. Like for a business expense, it's a very reasonable expense. So use those tools instead of you spending and wasting time. Your hourly rate is comparatively so high that it's way more worth it for the business to spend it the one times a month thing than have someone just work on that.

30:43So I'm gonna let this run in the background, But does this answer your question of how to do this for yourself? Yes, perfect. Thank you so much. Wonderful. You had Fritz, you had Atharva, I think, and Eric. What are other things, other routines that you would like me to break down and show you how you could create for yourself? What about routines where it has to maybe access a webpage? Okay. Where it requires a login. Let's say it needs to go into Eventbrite and book an event for you. So what would you say is your objective here? You're wanting to sign up for events that are relevant? Yeah, so I built an agent that daily goes through and find relevant events for me. And then right now it is sending it to me via Telegram.

31:45However, I wanted it to automatically book as well. And I was finding that a little challenging. Okay, wonderful. So step one is you've already got this figured out, but I want to write it out for everybody who's watching this or listening to this. So find relevant events. What's funny is we actually already did this, like list out the people I want to track and find relevant events are actually very similar. They are, you give it a list of targets, a list of sources, event, bright, luma, et cetera, and you say, here are where the events exist, go find them and do this on a weekly basis, right? Then your question was like, summarize and create a list, plus create a list of them. And then you send me them for approval.

32:38which ones I'd like to attend. And so you've got this working already. And the last step that you want is for those I accept, book it for me using my account. Correct? Wonderful. In this context, the best way to do this is to use a headless browser that will take into account your OAuth, your user account, and act and impersonate you. That allows, it's because we need a way that you need to delegate that authority to your agents and Claude and chat to PT don't want to take that responsibility on their shoulders because it would be very easy for someone to blame them for saying, hey, you did stuff without my consent. And so you need to take ownership of that. And the best way to do this is to use a headless browser. Now a headless browser, what is it? There's a few companies that do this very well. The two I recommend to people are Browser Base and Kernel .sh. Now Browser Base is right here, Browser Base, dah, dah, dah, Browser Base right here. So BrowserBates is a...

34:04software service, and I think they have a free tier that's very reasonable. So three concurrent browsers, one browser hour, three agents included, thousand search calls. So this would actually, that this would absolutely fit what you're trying to do. And you would say for the last step of the routine, use this connector, use this tool to go do this for me. Right. So what happens if I go to Kernel SH, because I do believe I've got my, I've got an account with them already set up right here. I can launch a browser here, and you see it's going to open up this new browser like this. I could absolutely go to something like click to take control, just so you show it like Luma. Let's just actually just search it, Luma.

34:54Then I have Luma right here, Luma, Luma, Luma, Luma, sign in Luma. There you go. Now, I can go here and I can say, okay, sign in with Google, sign in with Passkey, I could absolutely go ahead and sign in with Google, and I could go through and authorize this. What this will allow is actually, on this side, you actually have here managed auth. So managed auth is actually allowing you to give it connections, credentials, and integrations. So I could go ahead here and add that specific connector, that specific account. So I could give it my Luma account, I could give it my Eventbrite account, and I would give it all the accesses it needs to. And this kind of actually leads back to Atharva's question initially, which is like, how do I keep all these things secure and how do I delegate access? This is the way you do it. You find services that have specialized in this and you use their secure vaults to do this Because what you're doing is you're separating, you're reducing your blast radius at that point.

35:59Blast radius is a terminology used by software developers to say, if this goes catastrophically wrong, how far and how big of an impact does it have on me? And so our objective is always like, the worst thing would be delete my computer. Like that is the worst blast radius is it goes into your computers on a Windows, it takes this to 32 and it just deletes it. That would be catastrophic. You want to reduce that as much as possible, and the way you do that is by keeping your accounts secured within these auths. So there you go. That's how I would approach it if you just want to make this seamless for you. Wonderful. I'm curious to see here. So Apify can't be reached, that's not a problem. So I'd have to connect Apify at this point. Use Apollo, just use Apollo auto.

36:54This is just like general purpose. This is, I do wanna get the 100 people here so that I can show Valerie. That'd be really cool. Okay, next one. This would be a complete one at this point. I would just add a do a summary and log actions in a spreadsheet. good habit when you're building these kinds of like things that are doing things on your behalf is have them log what they do in some form of spreadsheet or database so you have like an easy way to go through the history without needing to click on like 10 ,000 different applications because like agents they are digital natives right they live in this digital space for them going to 10 different applications is really easy for me as a human it's It's a headache, and it makes my life horrible and miserable. So have them log their actions in a database when you do this is a good tip.

37:53Wonderful. Another workflow. I love these. We have 10 more minutes, so I could probably do two more, I think. I don't have anything super specific, but honestly, hearing some of these things is really useful for me. I'm really good at big, weird, complicated stuff, implementing code -level gates and that type of thing. I'm really bad at, hey, I just want to automate this simple thing and like get the stuff off a web page and everything's like, oh, you could use this headless browser and the service here. And there are things that, you know, a lot of people pick up and I don't know them. And sometimes I wish I did. I will, I will teach you all the tools there. No problem. I think tools are, tools are really good. I think people get a tool obsessed in like maybe they start pursuing to tools as as the as the end and not the means.

38:55But yeah, tools are always a very good way to get what you want. And it's always more efficient. Yeah, it seems like it. And again, that's where I make a bunch of tools for automating huge things with AI processes, like my personal database, I think is up to 8 billion records now, and all kinds of real -time aggregations, automated backups, hundreds of structured reviews, all scored and eval, that I can do. How do I get something from a website when I just need something? I don't know, I got to go look. Yes, I'm missing that part. Yeah, no, I get that. I get that. One of the the top three automations you should have in place.

39:49So these are routines that you should probably have active on your own workspaces. The first is like, these are routines. It's the same process that I just showed you. This is exactly what to do. This is how to do it. The first is what I call a weekly review. The weekly review should go through all of my chats this week and extract what I've done. We're constantly doing new chats and new windows and new things. The best thing to do for yourself is to keep tabs on that for yourself. which is, and you can make this weekly review become much more of a behemoth if you wanted to, right? Like right now, I said, go through all of my chats, right? But I could also add here emails, transcripts, meetings booked in calendar, right? You can actually go ahead and this is the promised Jarvis.

41:02The problem is Jarvis from all of these companies exists already. The thing is it does require a tiny bit of setup, which is the reason it requires a tiny bit of setup is because we're all different and we all use different tools. And therefore, like you can't just have this like one size fits all automation. You need to kind of build it for yourself and adjust it based on what you care about, right? So this is the biggest unlock for anybody is have an automation that goes through your week and gives you a recap of it. I'm not saying make advice. It's not making recommendations. It's not telling me what to improve. It's not making a judgment on what I've done. What it is doing is data entry. It's going through everything I've done in my week and it's saying, here's what you've done.

41:59Can I ask you, Simon, what's the purpose of that? I mean, you've had a plan, and then you're preparing a plan, but why do you want to do that? Why do I want you to do this? If I do it this week, it's not very useful, because this is week one. The thing is, when I'm at week 52, I have no recollection of what happened on week one and I need a I need a quantitative way to measure what have I accomplished what did I do this month what did I do this week what did I do this quarter and how does it compare to the previous period because I need to understand my own growth I need to understand what I've been doing has my volume of work increased has the nature sure of the work I've done changed. And so the usefulness of this is not in the short term, but in the long term. If you set up this automation, you're starting to create this ledger, this source of truth that you can turn to to understand how you've grown and how your work has changed over time.

43:19Now, this is number one. Weekly review, super useful, super easy automation. Next idea is filter out content and give me summaries. So I'm going to go and get an example of this on another browser tab, because I have one that comes in every single morning of this, but I do parenthesis here. Every morning is oftentimes too much. It's way too much to receive stuff every single morning. just a second, go through my emails right here. There you go. So here is something that I receive from my agent every single morning. This is a brief.

44:07Here you go. Copy, paste. So I'm not I'm not going to go into too much depth in here, but what I'm going to do is right now, this is a content agent that has an email I'm using instinct for this right now, but you can use cloud you can use anything. You can use any model, it goes ahead and it scans, all of the things that launched in the AI space over the last 24 hours, seven days, etc. So it's like, okay, I want to go and check out what is the new news. So I see that XAI launched Croc Voice 3, I see that Cohere did these things, I see that Antropic released these press releases and alignments and Meta. Meta did this, OpenAI did this. So this is like the second thing is instead of having a news, paying for news from multiple sources, which you can keep doing, If a subject is interesting to you, create a routine to go and get that content and have the AIs find these pieces of information for you.

45:15Again, this has two purposes, two benefits. The first is if it interests me, I can go read more. And two, it's over time I create a log of all the things, which allows me then to create what I find even more useful from this, which is trend analysis. So when you start doing these like scrapers and crawlers that go ahead and find content and make summaries, you can actually start figuring out trends of like, where's the market going, right? For example, security is becoming like a big issue right now. And because all of my articles are actually swaying towards alignment and they're swaying towards all of these like really important topics. And so agent observability and cost control are becoming a first class thing right now in the market. Voice is a live benchmark. There's a lot of competition going on in terms of voice right now.

46:22You have use, you have open AI, you have rock now. Like there's a lot of competition in that space. And so, well, me as an AI implementer, I'm like, oh, this is interesting. What that indicates is within a few months, because these are the model providers, right? Within a few months, there's gonna be big improvements in terms of voice AI and voice agents in what's going to be happening in the in the in the coming weeks and months so now if a client asks me what are some trends what are some things happening well I can tell that security is really important agent observability and cost is the big thing and voice is like the next frontier in terms of consumer tech right and so this is why I would do this. I know we have almost no time left so I'm going to do, just give you quick things, the third one. The third one is help me, da, da, da. OK.

47:16The third automation is not applicable to everybody, and that's a bit on me, but it's like help me generate images. I am someone who is extremely graphical in nature. I am a visual learner. I care about learning things visually. And whenever there's a concept that I don't understand, I constantly go to like chat or Claude, and I ask it to generate images for me to explain concepts. Instead of trying to give me text summaries, which are very exhausting and lead to my mental state not being very good, I ask it to create images or comics or whatever format is applicable for me to learn best, right? And that to me is images.

48:08So if there's one thing that you can do, like a routine, is have it do this like same kind of content filtering or weekly review, but have it make it graphically for you. Make yourself your own webcomic of your own life every single week. And it'll probably help you both retain more stuff, but also there's this like side effect of when AI starts being more proactive and sending you stuff, it makes you feel better about things. It's easier to interact with AI. You feel good about it. So help me generate images or visualizations visualizations of my week concepts on a weekly basis is something I strongly recommend to everybody put in place. It's a quirky way of using AI, in essence.

49:09Wonderful. I know we're at time. Is there anything else that came up? Any comments, any things that you guys would like me, like subjects that you would absolutely like me to cover? Next week, I had planned lead lists, how to generate lead lists, how to generate like prospects for yourself. And by the way, Valerie, while we were working here, I will send this to you, but here are a bunch of leaders from the big four, PwC, KPMG, EY, Deloitte. And you can use this as a base for yourself. So then you can actually go ahead and create that automation we talked about. And you can, here, let me just download this as an Excel and I will send that your way. Mind -blowing. Thank you so much. This is very cool. Wonderful.

49:57So, is there any topics that you guys would like me to cover in future sessions knowing that next week is lead lists? Go ahead Eric. Sorry, go ahead. Oh, have enough. Okay, have enough. Go ahead. A quick question for when you're doing, you're using a lot of agents to develop software, how are you managing your CI because the CI Bill starts to wake up? Maybe not connected to this, but yeah, that's something I am dealing with. Okay, I'll send you a more in depth like one -on -one answer with regards to that. To summarize to everybody, CI is just like the development process, continuous improvement.

50:50You'd want to find the balance between you're going to want to spend a bit more time on the planning phase. Generally speaking, if you're doing your continuous improvement is like that, it means that your agents will need a bit more coordination, and you're going to want to make bigger pushes and bigger, you're going to want to do offload a lot to smaller models and then do bigger pushes is, is the first impression I have. Okay, thanks. Wonderful. Awesome. Perfect. Eric, did you want to say anything or you're good? I think I'm good. Honestly, it's really helpful. I appreciate this. The top three automations are, are super useful. And honestly, the lead list for tonight really should learn as well. So looking forward to that.

51:47Wonderful. Take it away from me, Ms. Tussbrough. Good. If everyone wants, there's luma .com AI in context. I've actually set it up so that you've got all of the next ones that I've thought up. I will be changing these because I don't think proposal generator has not garnered much interest. If you guys want to set up, if you guys want to subscribe to a specific thing that is more interesting to you, go ahead and do this. I'm going to try and update these. I'm going to really try and do these every single week other than like your major holidays. I'm going to be taking Christmas off for sure, but just all that to say I intend to do this on the long run and try to give away all the secrets. I think there's a lot to be learned and just sharing what we've learned and making this successful to as many people as possible. So, wonderful. Well, thank you everyone for joining and hope to see you next week. If you you can, don't worry, sign up anyways, you'll get the recording. And I will let you guys know and send you guys an email and make this work. My agent Swarm is ready to take this Zoom recording and upload it to YouTube and send you guys emails and do all the things, so.

53:03Thank you, Simon. Thanks so much, Simon. Perfect, thank you. Thanks, everyone. Bye -bye.