How to level up your productivity with MAS (Multi Agent Systems)

WhatTheStack 2025 / 33:28

Watch this talk on YouTube

Transcript

32 paragraphs

This is an automatic transcript of the recording above. It is published in full and unedited, apart from correcting names the recogniser reliably mishears. It will contain mistakes.

00:09[music] Thank you so much. Wow. It's so good to see you all this morning. It's so awesome to be outside. You don't have to like smell a room. I can smell nature. Uh it's good to be with you. I'm I'm Tagious. Uh that's pronounced like contagious. Don't worry. I'm not contagious. It's just a it's just an easy way to remember how to pronounce this thing. And I am an engineer at IBM. I I work on AI and I work on developer relations for AI. In other words, I help developers like you have a good relationship with with IBM with our products and with AI, right? And my goal here today is to talk to you about multi-agent systems without any hype, without any Can I swear here? No. Yes. They're like at without any [ __ ] Okay. We'll we'll have a real look at AI and agents and multi-agent systems. To get started, if we want to talk about multi- aent systems, we need to understand, okay, but like what's an agent? Um, people started conversations this year by saying things like 2025 is the year of agents and I' I've heard this many times and I've thought to myself, who starts a conversation this way? What about what about just saying hi? You know, 2025. And so, um, if 2025 is the year of agents, what does that mean? What is an agent? Let's let's start from first principles and kind of build up. There's a lot of impressive demos today, but if we don't have a shared understanding of what an agent

01:34is, we're not going to leave as impacted. Okay. So, what is an agent? An agent is just one with agency. That's what it means. An agent is someone or something that has agency. Okay. But then what is agency? And agency is, you know, it can be a company that does design. You have design agencies, right? That's not what we're talking about. Agency in terms of personal agency u is something different. It's as defined by the Miriam Webster English dictionary as a person or thing through which power is exerted or an end is achieved. Right? And so when a person can do something, we can exercise power to do something and achieve an outcome that's called agency. Example, I woke up this morning and I decided I'm going to give a talk.

02:23So I came here, I used my agency to come here. Does that make sense? That's what agency is. And an agent is one who has agency. A human agent is me is all of us. We have agency. We in fact you chose your chair with your agency. Um but AI agency is exactly the same. It's when a machine when an artificial intelligence can exercise agency. Okay, what does that look like? Well, if you ask me to multiply 18 prime numbers in a row, I would tell you I don't even know what prime numbers are. I'm joking. I I I know maybe what the first few are, but I don't know like, you know, past five of them. I have no idea, right? Um thankfully, I know that there's a tool a tool exists that can help me with this. What is this tool called? A calculator, right? And so if you tell me, hey, multiply the first 10 prime numbers, my language model, my brain will tell me I can't do that. But there's a tool whose description is use this for calculation. So then I go use the tool and follow what happens here.

03:24My language model, my neural network, my brain creates the inputs to the tool. Multiply three and seven. So three comes from my brain. I press the button three, right? And times seven. So the inputs come from me. I press equals. The outputs from the tool come back to me. Does that make sense? That's a loop. And I'm using my agency to recognize when I don't have the skill and I use a tool. This is copied one to one with AI agents. Exactly that. But artificial is what we're talking about. In fact, let me show you a demo of an of an agent. Um, will my phone slide off? Okay, let me show you a demo. So this is Check this out. We're going to open the calculator. Check it out. My agency. I know this tool exists. Multiply uh prime numbers.

04:09The input three comes from my brain. From my brain from my brain equals And now I get the value and I like you know it comes back to me. Um I don't know if three is a prime. I'm really bad at math. Is it? Hey, thanks. Awesome. So that's kind of a stupid demo to be fair. Let's like do a real demo. Um this time of of what I just showed you but with AI. Okay. So uh to do this I'm going to use a tool that we build. It's called Langlow. It's open source. It's a free desktop open- source thing. Uh you can just download and use it yourself if you want. I'm using this because I believe it's the best way to visually build AI agents. You will understand together how this works. So I'm going to open up this is Langflow right here. Uh and I'm going to just blank canvas. And what we get is an actual let me connect to my hotspot uh first. This is going to be uncomfortable cuz now there's like you just have to watch me try and connect to stuff. Come on.

05:08Somewhere it's going to you're going to see Tis's iPhone. Please nobody connect to that. It's kind of important for me. Uh what' you say? You have to turn it off and on. There. There it is. T's iPhone 16. Fantastic. I love the ecosystem. You know, I really don't like Apple these days if I'm being honest, but the ecosystem. Chef's kiss. Okay. So, let's build a basic agent. An agent you talk to, right? Even your brain, there's inputs and outputs. So, we'll do a chat input and a chat output like this. And I'll just drop in an agent here. Just like that. So, now we've got an agent and chat input and chat output. And we drag the input. We drag the output. Right? And now, if we test it, let's go to the playground and say hi. Um, we send some input and we get some output.

05:53It seems like you've provided whatever whatever. So, it's working. It's working. We're talking to GPD4 mini from OpenAI. It's working. But is this an agent? Kind of. But it has no use for its agency. It's just like giving me language. This is the same as like Chad GPT in its basic form. How we can build an agent is we give it tools to use for the right job. So let's try again. Let's this time let's add a tool. Let's add a calculator. It's exactly the same analogy. This time I'm adding a calculator and I'm turning on tool mode. Okay, that contrast is quite bad. Let's check if dark mode looks better. Uh dark dark mode. O, can you see that clearly or is white? No. Okay, you like white. Okay, so we have a calculator now and I'm going to drag this tool output here and just like that it can do math but accurately. So I'll start a new chat multiply the first 10 forced the first 10 prime numbers, right? Um and now watch what happens. accessing that's it accessed a tool the product of the first 10 prime numbers is this and I can see okay so you used evaluate expression which is a tool from the calendar and these are prime numbers the first 10 and so it used the tool based on the con is that make sense and so you might think okay that's cool that's what an agent is but it's kind of basic um what if we could increase our productivity I suck I'm really bad at maintaining and managing my own calendar I'm really bad

07:20at um ask any organizer of the conference. Uh and so I need an agent for this. I need help with this. What if we could have AI just manage our calendars for us? Let's let's look at that. So, um I'm going to let's leave the calculator, but I want to give this agent now another tool. It's called calendar. It literally will connect to my Google calendar and do work. And I can when I turn on tool mode, I can choose what capabilities it has. In fact, I'm going to say you you can't do anything, but I want you to be able to create events, delete events, list events, move events, right? So, you can say I want you to be able to do these things. And I'm going to connect the tool set. And that's literally it. I have now an agent that can. And to show you how ridiculous this is, let me open this is my calendar today. And I'm just going to split screen this. So, we've got my calendar and we've got my agent.

08:10Can I hide the sidebar? There we go. Okay. Okay. And so now I'm going to open up this chat. Um, we'll start a new chat and it'll be like add a calendar event to speak at WTS conference at 10:25 a.m. Europe Berlin time for 35 minutes, right? Um, and I'll just send this off and you can watch in real time the the agent go to work somewhere. Actually, it's it's kind of the there was There you go. So, was added. Did you see that? This this didn't exist before. It just did the job. That's ridiculous. And if we go look at the chat, uh I believe it was here. Um no, it was this one. Yeah. So, it says Oh, no. It didn't actually update the Let me just reload this window. Um the playground has some stale state. So, let's Yeah. So, here we have the I added a calendar event to your calendar and it did. It exists here.

09:06That's incredible. It It does the job. It should. It can. And now I can, you know, I can even be like, "Okay, cool. Now add lunch at 1 p.m., right?" And I don't need to include more context because it kind of knows the time zone. It knows everything. And so it's just going to, if you look over here, um, there we go. Lunch just exists now. Um, and so I have this thing that can manipulate, it can do work for me in real time. Incredible. I deleted those events. But that's a single agent system and it it can operate my calendar fairly well. But we're here to talk today about that multi- agencies. Why would you need a multi- aent system? To recap, we kind of understand now agency, what agency is, the ability to know when is the right tool for the job and use it. We know what an agent is as one who has agency. And finally, we can talk about multi- aent systems. Okay, you might be thinking, cool, that calendar agent was great. Um, but I can just add like many tools to it, right? Why do I need a multi- aent system? You may be literally like why is this important? The reason I will say multi- aent systems are important is the same reason um companies exist. If you think about a company where people work, a company is really just a multi- aent system where there's a bunch of human agents. If you think about civilization, if you think about your government, right? Like they're all multi- aent but like real

10:25human agent systems. And you need a multi- aent software system for the same reason because not one person can do every job. right? If we could, we would uh but it's not the case. Uh and so you need a team of specialists. You need like a front-end expert and a code reviewer and this type of thing. That's the gist of it. But if you insist on more reasons, reason number one is cost. There's a real irony here because a lot of companies want to fire people to save money, right? Uh more people, more cost, usually more agents, more human agents, more cost. Um usually with AI, it's actually counterintuitive. With a multi- aent system, it can actually be cheaper than more expensive. And the reason for this is not all queries from a user require like a big model to be used. You don't need like a 60 billion parameter like GPD4. Some queries are so simple you could use a very basic model and get a reasonable output, right? So your cost goes down. In fact, a a very classic multi-agent system is one that just saves cost by routing complex queries to a more expensive model and simple queries to a cheaper model. Number two, security. Um, if you think about the internals of your computer and your phone, it works like a multi- aent system. There's a lot of delegation.

11:41What do I mean by that? Well, you've got a CPU for a type of processing. You've got a GPU for graphics processing. And on most devices, you've got a secure enclave, it's called, where things like your fingerprint for Touch ID or your Face ID, that data is stored in there. And it's it's tamperproof. Nobody can get that data. In fact, whenever the CPU or the OS wants to access it, it has to give a special cipher. It talks only in encrypted values, right? And so there's a multi-agent system. There's an agent for secured data. There's an agent for graphical data and an agent for regular data, right? A CPU is like that. So with security you can have one agent to handle the secure internal sensitive stuff and an agent for broader. This was actually the original design of the failed Apple intelligence. You know Apple intelligence on the iPhone would use a local ondevice model for your phone numbers, your contact information, your photos, your personal stuff and then when it reached the end of its limits it would call out to chat GPT.

12:41This was the it was a multi- a intelligence was a multi- aent. Does that make sense? So security is a is a reason for a multi- aent system. Finally, analysis paralysis. If you have many many tools, there's a chance you're not going to use any of them, right? And we'll talk about that in a little bit more detail. But in terms of cost, I wanted to show you I did an experiment. Part of my job is to do research. This is the cost between a large model like GPD4 and a smaller open-source model like Mistl 87B. And so you can see that blue line is ridiculous, right? the costs go up to like $20,000 per um million prompts and the the red line is obviously it's like so much smaller and this is a great case for a multi- aent system. If you want to save costs, you do something like that. If we zoom in a little bit, this is what that's ridiculous the amount of costs you save with a multi- aent system. The next thing I want to talk to you about is analysis paralysis. Like for us as humans, let's do an experiment. You go to a restaurant and there's like 300 items on the menu and they all look amazing. How do you respond to this?

13:45You're like, "Oh man, I don't know what to get. Do I get all of them? I can't get all of them cuz then it's going to be expensive and I'm going to maybe gain some weight and all like you struggle, right?" Um there was a paper. This paper is titled this. It's titled when choice is demotivating. Can one desire too much of a good thing? It's foundational and it proves that when we have more options, we take less action. More choice, less action, right? Um, this paper was actually cool cuz it was a paper on jams and essay topics and chocolates. And there were two groups. Group A had to choose one from six jams or six topics of study or six chocolate. Just look, you have six options. Choose one. Group two had to choose from about 30. Okay? And what they found every single time without much variance is when you have more options, you don't move. You don't do stuff. You get paralyzed. Now, multi- aent systems or AI is like literally just modeled after our human brain. They're like us, you know, like if we know how to use a tool, they know how to use a tool. If we struggle with analysis paralysis, they struggle with analysis paralysis. If we are discriminatory and racist, they are.

14:55Anyway, um the it's true and that's also why you need a multi- aent system because you can like correct for inappropriate behavior. But this is a multi- aent system. It's very possible to give a system many tools and then it does nothing. In fact, I'd like to show you that now. We're going to take our calendar example and make it multi- aent. Okay? And let's let's just open the calendar right here. We don't need this one anymore. So, this is just a single agent calendar that we built. I'm going to get rid of the calculator, but I'm going to introduce complexity by giving not one calendar, but two calendars. And you watch this thing just like freak out. So, I built my own tool to interact with Apple calendar. So, I'm on Mac OS and this is my calendar app right here. This is just the default like built-in Mac OS calendar. And so, I'm going to go use my calendar tool right here. I'm turning it into tool mode. And now, by the way, all these components here in this graph, they're just code. Like if I click on code, you literally just write code and it becomes a component. Anyway, so um I'm going to give this Apple calendar. So now look, we have two calendars. We have Google calendar, we have Apple calendar. Um what's going to happen, right? So if I go back to my playground, I will copy exactly the same prompt. Add a calendar event to speak at WTS conference. And I'll send this prompt again. And notice

16:12I I want to draw attention to this. My Google calendar is is empty today because it's Saturday. And my Apple calendar is also totally empty. There there's nothing on there. So now if I send the same prompt, um, what's going to happen? Let's find out. So it has more choices. Cool. You're getting the current date and you're adding it to but which calendar? Notice it doesn't specify. I added the calendar. Cool. But let's go check Google. It's empty. Apple, it's okay. It's there. But like I didn't ask for Apple. just kind of choke. This is the problem right now. If somebody asked you if if if they had two calendars and they said add this to my calendar as a human agent, what would you do? Would you add it to both? Would you ask them a follow-up? Like which one? You know what would? And this is a real problem. This is why we need multi- aent systems. So how we can fix this is by turning the single agent system into a specialized multi- aent system. And we can do that very easily with a little bit of effort here. First thing we need is we're going to keep our chat input and we're going to duplicate our agent.

17:17So we have two agents now that are kind of the same. We'll delete the output and we'll have one agent for Google and one agent for Apple. Literally just like different assistants with different use cases, right? And so we'll say this is the Apple one and this is the Google one. But we need some way now to balance which agent do I talk to? This is called the orchestrator pattern. We don't need two agents. We need three agents. We need one agent to receive the query and be like what does the user want and then when it knows what the user wants route to the right agent. Okay, this is this is the most common pattern with multi- aent systems. So how can we do that? Well, we need a we need an orchestrator agent. So I'm going to duplicate that.

17:57Bring it here. And my user's query goes to the orchestrator agent. The system prompt the identity statement for the orchestrator agent is this. I don't I hope you can read that. I'm I'm going to narrate anyway. It says this. You work with the users's calendar, both Apple and Google. You add events events to either one or the other depending on the user's query or both. Right? So, you're just giving some identity here. Cool. That's the system prompt. So, now how does this agent use these other agents? You can turn an agent into a tool for an orchestrator agent. So, I'm just going to turn on tool mode here for agent, the Google agent. Turn on tool mode for the Apple agent. And when I turn on tool mode, um, I get some tools here. And I have to now do the work of describing what each tool does. The only reason as a human agent I know when to use a calculator is because somewhere in my life, someone described a calculator to me. They said, "This is a thing that you use to do number stuff. And then I know we need to do exactly that because they speak English. They speak language. So let's do that. So now we have the Google agent and we get a tool. So look at this. This is super generic, right? It's just it's called call agent and a helpful. This this means nothing. Our orchestrator is not going to know what this is for. So we'll fix this. We'll call the slug use Google calendar. And

19:26the description is use this tool to work with the user's Google calendar. Right. Um and we'll get rid of that. And we'll do the same for the Apple calendar. We'll we we have tools now and we'll just describe them. We'll change this to Apple calendar and we'll say use Apple calendar. That's the name of the tool that this agent is. Okay. And finally, we take these two agents and we connect them as tools to our orchestrator. You see, it's a multi- agent. There's three agents now. It's fantastic. All right. Let's let's use this um let's use it. So, what I'm going to do is open up the the playground. My goodness. I I've zoomed too much. How do I change? I I I've lost my sidebar because I just zoom zoomed so Okay, we're back. So now I'm going to open the playground and I'm going to send exactly the same prompt.

20:18Where is it? Here. Add a calendar event to speak. And now I wonder what it'll do. Again, there's nothing in my Google calendar. There's nothing in my Apple calendar. So let me add a new prompt and say this. And now we're working with a multi- aent system, right? So what what's going to happen? Oh, actually nothing because there's no chat output. This guy, this agent is not talking to a chat output. So let's go back and stop this and let's drag a chat output right here. So we can actually see what's happening. Did it touch the calendars? It did not. Okay, great. Let's go back and send that prompt this time with a chat output and we'll see what what happens. So we it's it's processing our input and it's using Apple calendar. Okay, so it just chose Apple by default. I I wonder why. Um, okay. So, it added it to the Apple calendar. Look, the event speak at WTS has successfully been added to both your Google calendar and your Apple calendar for October 25th. October 25th.

21:16I have no idea why. Um, but let's check. So, Google calendar. Let's go to October 25th. Some hallucination there, wasn't it? Uh, October 25th. It hasn't even been Okay, this is bad. Let's AI. Am I right? Um, can I view it in my Google calendar? Did it do it? Oh, cool. It did it. Yeah. October 25th, 2023. Um, okay. It did it anyway. And let's check the Apple calendar. Let's go back. October 25th, 2023. October 25th. Are you there? Okay, it's there. Cool. It did it. Um, it did it. I didn't, you know, give it it. I could include the context like the date like it is 2025 whatever but it did the job and it worked with both calendars. Let's tweak this and say just do one right um add to speak at WTS conference on Sept September 2025 and we'll say add it only to the Apple calendar. Okay. And we'll send this off.

22:16And so we'll go back to Google calendar. We'll go back to today. Um, today and we'll go to the Apple calendar. We'll go back to today and here speak at this was just added but it's not on the Google calendar. And if we look at the log, thankfully there's logs. What we can see is it only executed use Apple calendar. It did not execute use Google. In fact, just for completion sake, let's run the same prompt with this time only to the Google calendar. And we'll delete the Apple event, right? Um and now what we should see is it will speak only to the Google. So we have this cool orchestrator mult. Yeah, it's using the Google calendar. Fantastic. Um do your job. And if we come here there speak at WTS. So we have this cool orchestrator but you might be thinking great but like why and I hope it's not lost on you that we're not this is not about calendars. This is about when to let an agent use its agency because instead of Apple and we could we could change the variables and say that my Apple calendar is high security. It's high security. Nobody's allowed to it's not connected to the internet. Open AAI cannot know about my Apple calendar, right? We can say my Google calendar because it's with Google and Google spies on all of us anyway. Um that's public, right? And so we we get more rules like this with a multi- aent system. Let's take a look at how we can work with security. Again, the three

23:38reasons for a multi- aent system. Number one, cost. Number two, security. Number three was analysis paralysis. So, we just solved for the analysis paralysis case. Let's solve for security and cost by saying if it's Apple, then only use an open-source model. Do not talk to OpenAI, but keep everything on this device. Zero internet dependency. We can do that with a multi- aent system. Let's let's look at how we can do that. So, we'll come back. This is working as expected. I don't have any calendar events. I've deleted all of them. Clean slate. Let's now say for the Apple agent here, this is my Apple agent. Um, we don't want to trust OpenAI. We don't want to trust Google Calendar. We do it all on this MacBook alone. So, I'm going to say instead of OpenAI, let's use a different model. And there's a tool called Olama which will allow you to run any open- source model on your MacBook.

24:34It's it's open source. Um and so what we can do is say use Olama and Olama listens on local host localhost port 11434. That's llama in lead. Okay. And you can choose a model GPTOSS. This model here is OpenAI's open source model. So it's like using OpenAI but without the spying. Okay. With privacy. So we'll do GPD OSS um and that's it. That's all it is. So now the Apple one is going to use Olama locally on device and it's going to talk to my ondevice Apple calendar. It's the secure one and the Google one we can assume is more public. In fact we can change the system prompt instead of being so clear about Apple and Google here. We can say this you work with the user's calendar both a secure calendar for sensitive information and a public calendar. Okay.

25:25You add events to either one depending on the so we changed the scope. This is no longer Apple or Google. It's like security or not security. Also cost goes down here because we're using an open source model for free on our hardware. Okay. So now we just adjust the tools a little bit saying use this tool to work with the users less less sensitive public calendar. Right? And we'll change this to say use this tool to work with a user's more sensitive secure calendar. Right? So now we have security built in. So let's try let's try let's try let's try let's try. So both my calendars are clean. There is nothing in them. Let's try and say the same thing here. Um add a calendar event to speak at WTS conference um for 35 minutes. I don't even know which calendar it's going to choose. That's the cool part. So, let's just see. And what is it choosing? What is the So, it's using the Apple one because it thinks this is more of a secure thing, right? And it's probably only using the Apple one. So, let's uh and and you notice it's slower cuz it's on my device. Take a look. This is my GPU usage 99% because it's running a model on my Mac. In fact, when the GPU usage comes down, then we know the the job is ready. Okay. In fact, let's open this side by side and we can see when the language model stops working. And again, it's local. It's secure. But there, okay, so look at this GPU usage 29%.

26:52It's done. You see, um, and it added it to both my public and secure calendars. So there's it's in Apple. Great. And it should be in Google as well. According to the AI, it is in Google. Fantastic. But what if it's something sensitive? Let's paste that prompt again. Let's come here and we'll say um let's okay at a calendar event not to speak at the conference but for the birth of my firstborn son. This seems like a more secure like personal thing, right? Um so let's see what it does now with its agency. Does it decide it's Apple only? Does So it's it this is apparently public knowledge and it's doing both. Um but this is where we kind of have to fine-tune the prompts and so on in ir irrespective of it. We have one secure local ondevice thing that uses my GPU and one more public one in this multi- aent system. Um hopefully by now we understand multi- aent systems what we can use them for. The the Olama local model is going to cost way less money.

27:55It's also going to be way more secure. nothing leaves my I could have the most messed up conversations with it of all all kinds of evil plans to take over the world and nobody will ever know, right? Um thankfully OpenAI will not know. That's one agent. The other agent is secure and you can now have a multi- aent system. This wouldn't be a a balanced look if we don't talk about trade-offs. So let's take a moment and talk about when to use what because that's also part of the talk description which probably helped you use your agency to come here. So trade-offs when to use what? Number one, you will never get as deep information as with a multi- aent system. If you use a single agent, single agent systems can only give you like one level of information, but imagine like a deep research use case.

28:40In fact, I have a deep research demo. I don't have time for it, but come see me after. Um, with a multi- aent system, you can have many different agents. One to like if the user gives you a query, Scopier for example, then you can have five agents, right? Agent number one, take the topic and give me four sub questions. Agent number two, research those sub questions and give me sources. Agent number three, go browse the internet and find, you know, like so you can have many agents to deeply research things. With a single agent system, you can't do that. You get one answer. Okay, so multi-agent systems win here. Number two, there's this principle called aams razor which will tell you that the simplest solution is often the best solution and as engineers I think we agree, right? Like you if you can do more with less, then do it. Multi-agent systems are inherently more complex.

29:23Therefore, it's going to be hard to maintain and build. If you want simplicity, single agents will work better. Number three, wi when you want a specialized task, for example, judge this whether it's highly secure or public, if you want a specialized task like that, you need a multi- aent system. You need a similar to a company, you need a really good front-end engineer and a really good backend engineer and a really good like DevOps engineer and you can ship anything, right? But if you have a single if you just have a front-end engineer, you're not going to have tasks. If you have a full stack engineer even, you're not going to have task specialization. So this is another case for multi- aent systems. And finally, single agent systems win with latency, no doubt, right? Because you process not that much and you process quickly. And so single agent systems will get you faster responses every single time. Of course, you can use caching and stuff with multi-agent systems to get some balance. But as a general rule of thumb, if you if you honestly if you just walked in here and you're thinking when do I use a multi- aent system, sometimes abbreviated MAS, multi-agent system or single agent system, SAS, um the rule of thumb is this. You just if you can do it with a single agent system, then don't use an MAS. That's the rule. But if you have, for example, 18 different calendars in your calendar app, you

30:38probably need an MAS, right? So I I would say save it as a last resort. Okay, you might be wondering this is all cool page is great good stuff but like is this actually useful? Like how do I actually use this? Um, and I'd love to just show you that because lang flow that that that I built these with is actually something I use myself to actually like manage my workflows and increase my productivity because um this flow that we built this calendar managing thing um is not just a diagram but if I click on share and if I choose API access this whole thing is exposed over an API. So if I if I fetch this endpoint, I will invoke exactly that flow and so I can build any front end any mobile app any front end that's and you know I just type like add this to my calendar send a request to this endpoint and it just does the job and if I host this and add like a domain name tis.com/ai then this will just work like that right and it's already ready to go over the API. Um it also supports something called MCP or model context protocol where you can within your existing tools like cursor or windsurf. You can just use it. So it's the the the workflow is quite cool. You build a flow that works for you visually and immediately it becomes a web endpoint that you can connect any application to. Um so that's how you can use it. I want to start wrapping up with some next steps. Uh this was obviously a lot of information.

32:03I hope you got some some ideas here, some sparks of of inspiration to go build something that can help you manage your life, right? Some of you maybe here planning weddings and you're like, I don't want to deal with all this venue searching [ __ ] Thankfully, you can build an agent that does that. It browses the internet, finds the realistic options, etc. Okay, there's many applications. I hope I've given you some. If you want to follow up and continue, um, here are some things we can do. As you can tell, I love speaking about AI and about other things. Um, 35 minutes sometimes isn't enough. So I have a podcast [laughter] where where the discussions are much longer. There's it's a long form podcast of like 25 actually two and a half hours sometimes. Um and so if you're more interested in AI and where it's going without you know hype and clickbait with just like reason um I'd invite you to come listen to that. Um the the open-source tool that I used to build this um is called Langflow and that's something I work on as well every day.

32:57It's open source and it's it's a desktop application. You just download it and run it. uh if you wanted to learn more about lang flow it's that uh and finally it's not lost on me that many of you may have questions we don't have that much time for questions I am here throughout the day but if you want to talk about questions if you want to receive some slides um that's me on socials I'm more than happy to talk to you there as well with that I want to say thank you so much for your time and attention thank you

More talks

All 19 transcripts

Elsewhere

There is every talk I have given, all 69 of them, ConTejas Code, the podcast, and Fluent React, the O'Reilly book on how React works inside.