1 00:00:00,100 --> 00:00:01,966 welcome to money is freedom 2 00:00:01,966 --> 00:00:02,866 a podcast 3 00:00:02,866 --> 00:00:06,100 exploring how finance and freedom connect in our lives 4 00:00:07,000 --> 00:00:08,833 created from thoughtful research 5 00:00:08,833 --> 00:00:10,566 and narrated by Notebook L m 6 00:00:10,566 --> 00:00:13,266 a I this series brings you clear 7 00:00:13,266 --> 00:00:16,000 meaningful insights into finance and beyond 8 00:00:16,100 --> 00:00:19,633 welcome to episode 41 where we explore how AI 9 00:00:19,633 --> 00:00:22,266 agents are evolving from isolated tools 10 00:00:22,300 --> 00:00:24,800 into structured interoperable systems 11 00:00:24,900 --> 00:00:27,766 what some are calling the emerging agent stack 12 00:00:28,200 --> 00:00:30,600 this shift is reshaping the way we build and 13 00:00:30,600 --> 00:00:32,033 interact with AI 14 00:00:32,033 --> 00:00:33,700 and it's happening fast 15 00:00:34,433 --> 00:00:36,766 we're drawing inspiration from a fantastic 16 00:00:36,766 --> 00:00:38,766 blog post by John Turau 17 00:00:38,866 --> 00:00:42,866 the Agent Stack Emerges which maps out the core layers 18 00:00:42,966 --> 00:00:45,366 from memory and planning to orchestration 19 00:00:45,700 --> 00:00:49,133 that are forming the backbone of next gen AI systems 20 00:00:49,300 --> 00:00:51,100 if you're building in the AI space 21 00:00:51,100 --> 00:00:53,300 or just trying to make sense of the chaos 22 00:00:53,300 --> 00:00:56,200 this episode is for you let's dive in 23 00:00:56,200 --> 00:00:57,733 welcome to the deep dive 24 00:00:58,300 --> 00:00:59,366 today we're uh 25 00:00:59,366 --> 00:01:01,700 diving into something really moving fast 26 00:01:02,233 --> 00:01:04,166 the infrastructure behind AI agents 27 00:01:04,166 --> 00:01:06,100 yes it's evolving incredibly quickly 28 00:01:06,100 --> 00:01:08,466 we got some great material that shows how the actual 29 00:01:08,466 --> 00:01:10,466 you know foundations for these AI 30 00:01:10,466 --> 00:01:12,100 apps are being radically rethought 31 00:01:12,100 --> 00:01:14,766 not just built but redesigned exactly 32 00:01:14,766 --> 00:01:17,400 it's a real shift we're moving away from well 33 00:01:17,400 --> 00:01:19,400 what felt like a lot of one off experiments 34 00:01:19,400 --> 00:01:20,766 just a short while ago right 35 00:01:20,766 --> 00:01:23,133 and now now we're seeing um 36 00:01:23,133 --> 00:01:25,966 much more structure clearer layers are forming 37 00:01:26,000 --> 00:01:28,033 it's not just about speed increases either 38 00:01:28,033 --> 00:01:28,766 it's yeah 39 00:01:28,766 --> 00:01:31,600 it's making people fundamentally rethink 40 00:01:31,600 --> 00:01:33,100 software development itself 41 00:01:33,100 --> 00:01:34,233 that's really the core of it 42 00:01:34,233 --> 00:01:36,033 isn't it these foundational shifts 43 00:01:36,033 --> 00:01:38,666 yeah AI agents aren't just some SCI fi concept anymore 44 00:01:38,666 --> 00:01:41,300 they're actually pushing changes in infrastructure now 45 00:01:41,300 --> 00:01:42,433 right now so 46 00:01:42,433 --> 00:01:44,433 we're digging into this whole emerging AI 47 00:01:44,433 --> 00:01:46,100 agent infrastructure stack 48 00:01:46,300 --> 00:01:48,633 what we hope you'll get is a clearer picture 49 00:01:48,633 --> 00:01:50,400 of how fast this is moving 50 00:01:50,400 --> 00:01:51,966 and maybe surprisingly 51 00:01:51,966 --> 00:01:54,100 the patterns that are actually starting to stick 52 00:01:54,100 --> 00:01:56,266 and the growth speed is well 53 00:01:56,266 --> 00:01:57,833 it's pretty staggering honestly 54 00:01:57,833 --> 00:01:59,566 companies really leaning into AI 55 00:01:59,566 --> 00:02:02,766 agents are scaling in ways that just look different 56 00:02:02,766 --> 00:02:04,233 totally different I mean 57 00:02:04,233 --> 00:02:05,866 look at cursor they hit what 58 00:02:05,866 --> 00:02:07,966 a hundred million dollar a r r yeah 59 00:02:07,966 --> 00:02:11,433 in just 21 months with a tiny team like 20 people 60 00:02:11,433 --> 00:02:13,400 and then there's bolt and lovable right 61 00:02:13,400 --> 00:02:15,566 $10 million basically overnight 62 00:02:15,566 --> 00:02:17,633 couple months with 15 people each 63 00:02:17,633 --> 00:02:18,800 it really makes you question 64 00:02:18,800 --> 00:02:20,766 the old rules about scaling a business 65 00:02:20,766 --> 00:02:22,100 doesn't it absolutely 66 00:02:22,166 --> 00:02:23,466 is that sustainable 67 00:02:23,466 --> 00:02:27,233 or is it just a new kind of economic model emerging 68 00:02:27,233 --> 00:02:29,300 here that's the million dollar question 69 00:02:29,300 --> 00:02:32,566 maybe billion dollar but what it definitely shows is 70 00:02:32,566 --> 00:02:34,566 they're breaking free from old constraints 71 00:02:34,566 --> 00:02:35,900 it's not just you know 72 00:02:35,900 --> 00:02:38,400 shaving off a few percentage points of inefficiency 73 00:02:38,400 --> 00:02:39,566 no it feels bigger 74 00:02:39,566 --> 00:02:43,066 it's using AI for massive operational leverage 75 00:02:43,066 --> 00:02:46,100 and the infrastructure has to scramble to keep up 76 00:02:46,100 --> 00:02:48,033 to support this kind of hyper growth 77 00:02:48,033 --> 00:02:51,166 it's funny thinking back like less than a year ago 78 00:02:51,166 --> 00:02:52,700 June 2024 79 00:02:52,866 --> 00:02:54,600 people were saying this space was pretty empty 80 00:02:54,600 --> 00:02:56,366 oh yeah very DIY 81 00:02:56,366 --> 00:02:57,833 everyone building their own thing from scratch 82 00:02:57,833 --> 00:02:59,033 yeah except little labs 83 00:02:59,033 --> 00:03:00,800 exactly lots of experiments 84 00:03:00,800 --> 00:03:03,600 which is great but also loads of duplicated effort 85 00:03:03,600 --> 00:03:05,500 because there weren't really established ways of doing 86 00:03:05,500 --> 00:03:07,466 things and now boom 87 00:03:07,466 --> 00:03:10,000 just a few months later the pictures way clearer 88 00:03:10,000 --> 00:03:12,600 and we have actual data showing this speed up 89 00:03:12,600 --> 00:03:14,333 yeah like the neon report 90 00:03:14,433 --> 00:03:16,800 the serverless postgress folks right 91 00:03:16,900 --> 00:03:18,500 what did they find they're seeing 92 00:03:18,500 --> 00:03:22,966 AI agents creating databases on their platform 93 00:03:22,966 --> 00:03:25,766 more than four times faster than human developers 94 00:03:25,800 --> 00:03:27,500 four times that's huge 95 00:03:27,500 --> 00:03:28,566 it's not just a little bump 96 00:03:28,566 --> 00:03:31,233 it accelerates their core metric across the board 97 00:03:31,233 --> 00:03:33,100 what does that even mean for the future 98 00:03:33,133 --> 00:03:35,666 back ends becoming super dynamic maybe yeah 99 00:03:35,666 --> 00:03:38,300 maybe databases just spin up and down instantly 100 00:03:38,300 --> 00:03:41,300 based on AI needs it definitely points that way 101 00:03:41,300 --> 00:03:42,066 more fluid 102 00:03:42,066 --> 00:03:45,066 more responsive infrastructure driven by the AI itself 103 00:03:45,066 --> 00:03:46,566 and it's not just the back end stuff 104 00:03:46,566 --> 00:03:47,433 no definitely not 105 00:03:47,433 --> 00:03:49,100 look at browser bases stage hand library 106 00:03:49,100 --> 00:03:50,666 that's for UI automation 107 00:03:50,666 --> 00:03:52,966 letting agents actually use websites like we do 108 00:03:52,966 --> 00:03:54,033 ah yeah 109 00:03:54,033 --> 00:03:55,866 clicking buttons filling forms 110 00:03:55,866 --> 00:03:57,566 exactly and it just crossed 111 00:03:57,566 --> 00:04:00,666 half a million monthly installs on npm wow 112 00:04:00,666 --> 00:04:02,366 500 k installs a month 113 00:04:02,366 --> 00:04:05,700 that signals massive demand for letting AI 114 00:04:05,700 --> 00:04:07,266 interact with the visual web right 115 00:04:07,266 --> 00:04:10,266 absolutely for a specialized library like that 116 00:04:10,266 --> 00:04:12,966 it really underlines how intense the need is 117 00:04:12,966 --> 00:04:13,966 it almost makes you think 118 00:04:13,966 --> 00:04:15,366 if you're starting a company now 119 00:04:15,366 --> 00:04:17,366 are you already late huh 120 00:04:17,366 --> 00:04:19,800 you could think that but what's weird is we're 121 00:04:19,800 --> 00:04:22,900 still seeing brand new companies founded 122 00:04:22,900 --> 00:04:25,533 even this year getting real traction 123 00:04:25,766 --> 00:04:27,066 so why is that do you think 124 00:04:27,066 --> 00:04:27,466 well 125 00:04:27,466 --> 00:04:30,566 I guess it means the ground is still shifting so fast 126 00:04:30,566 --> 00:04:33,566 new needs new niches pop up constantly 127 00:04:33,600 --> 00:04:35,500 the core tech is still being figured out 128 00:04:35,500 --> 00:04:37,366 so there's room for fresh ideas 129 00:04:37,366 --> 00:04:37,566 so 130 00:04:37,566 --> 00:04:40,766 we're past the initial Wild West experimentation phase 131 00:04:40,766 --> 00:04:42,766 I think so that's a key difference 132 00:04:42,766 --> 00:04:45,000 a year ago it was mostly exploration 133 00:04:45,066 --> 00:04:46,633 developers starting from zero 134 00:04:46,633 --> 00:04:48,066 trying to invent the wheel right 135 00:04:48,066 --> 00:04:50,300 lots of trial and error figure out the basics 136 00:04:50,300 --> 00:04:53,266 but now we're seeing actual repeatable patterns emerge 137 00:04:53,266 --> 00:04:55,100 and smart teams aren't starting from scratch anymore 138 00:04:55,100 --> 00:04:56,366 they're using these patterns 139 00:04:56,366 --> 00:04:57,766 okay so what are these patterns 140 00:04:57,766 --> 00:04:59,166 can we categorize them yeah 141 00:04:59,166 --> 00:05:01,633 people are starting to identify a few common types 142 00:05:01,633 --> 00:05:04,866 first up there's what they're calling next gen copilots 143 00:05:04,900 --> 00:05:06,866 okay like the coding assistants we know 144 00:05:07,000 --> 00:05:08,900 kind of but more advanced 145 00:05:08,900 --> 00:05:10,300 they're context aware 146 00:05:10,300 --> 00:05:13,000 they proactively help with complex stuff 147 00:05:13,266 --> 00:05:14,800 think bolt dot nu 148 00:05:14,866 --> 00:05:16,600 air ops colimit 149 00:05:17,066 --> 00:05:18,600 how are they different from 150 00:05:18,600 --> 00:05:20,566 say git hub copilot 151 00:05:20,800 --> 00:05:22,166 well the idea seems to be 152 00:05:22,166 --> 00:05:24,933 they go beyond just reacting to your commands 153 00:05:25,000 --> 00:05:27,166 they actually understand the context of your work 154 00:05:27,166 --> 00:05:29,633 ah so they anticipate what you need 155 00:05:29,633 --> 00:05:30,466 yeah potentially 156 00:05:30,466 --> 00:05:33,300 offering help or suggestions before you even ask 157 00:05:33,500 --> 00:05:35,100 much more integrated interesting 158 00:05:35,100 --> 00:05:35,900 okay what's next 159 00:05:35,900 --> 00:05:37,300 then you get teammate agents 160 00:05:37,300 --> 00:05:39,033 companies like Ravenna sale plane 161 00:05:39,033 --> 00:05:40,900 base pilot are in this space 162 00:05:40,900 --> 00:05:43,666 teammate agents so more autonomous 163 00:05:43,666 --> 00:05:44,500 exactly they can 164 00:05:44,500 --> 00:05:46,866 actually drive multi step work falls on their own 165 00:05:47,000 --> 00:05:49,800 yeah it's like giving a complex task to an AI colleague 166 00:05:49,800 --> 00:05:51,500 who just runs with it okay 167 00:05:51,500 --> 00:05:53,866 so how does that differ from the next gen co pilot 168 00:05:53,866 --> 00:05:55,266 is it just more independence 169 00:05:55,266 --> 00:05:56,700 that seems to be the main thing 170 00:05:56,700 --> 00:05:58,700 co pilots assist you within your workflow 171 00:05:58,766 --> 00:06:00,666 teammate agents can own a whole process 172 00:06:00,666 --> 00:06:02,766 start defending with less direct input 173 00:06:02,766 --> 00:06:04,433 got it and there are more patterns 174 00:06:04,433 --> 00:06:06,966 yep we're also seeing agent organizations 175 00:06:07,366 --> 00:06:08,833 this is where you have multiple 176 00:06:08,833 --> 00:06:11,100 specialized agents working together 177 00:06:11,100 --> 00:06:14,100 like a team of AI's kind of ARRU's 178 00:06:14,100 --> 00:06:16,766 probabilistic simulations were mentioned as an example 179 00:06:16,900 --> 00:06:19,300 it's about coordination between different agents 180 00:06:19,300 --> 00:06:21,266 to tackle something really complex 181 00:06:21,766 --> 00:06:24,200 what's the benefit there is it just divide and conquer 182 00:06:24,200 --> 00:06:25,966 pretty much division of labor 183 00:06:25,966 --> 00:06:27,866 each agent is good at one thing 184 00:06:27,866 --> 00:06:28,900 and they collaborate 185 00:06:28,900 --> 00:06:30,600 almost like different departments in a company 186 00:06:30,600 --> 00:06:33,200 to solve problems too big for any single agent 187 00:06:33,200 --> 00:06:35,666 makes sense and the last pattern 188 00:06:36,000 --> 00:06:37,866 agents as a service basically 189 00:06:37,866 --> 00:06:40,200 companies offering specific AI 190 00:06:40,200 --> 00:06:43,000 agent skills as a service for other developers 191 00:06:43,000 --> 00:06:46,100 uh so like an API for a specific agent capability 192 00:06:46,100 --> 00:06:46,800 exactly 193 00:06:46,800 --> 00:06:49,400 instead of building say a web browsing agent yourself 194 00:06:49,466 --> 00:06:50,266 you just use one 195 00:06:50,266 --> 00:06:52,833 someone else provides plugging in expertise 196 00:06:52,833 --> 00:06:53,900 okay that lets developers 197 00:06:53,900 --> 00:06:56,266 focus on their main application logic right 198 00:06:56,266 --> 00:06:57,833 not reinventing every AI wheel 199 00:06:57,833 --> 00:07:00,066 precisely so all these patterns 200 00:07:00,066 --> 00:07:01,766 co pilots teammates 201 00:07:01,766 --> 00:07:03,533 organizations services 202 00:07:03,666 --> 00:07:04,366 they're obviously 203 00:07:04,366 --> 00:07:06,600 putting new demands on the underlying tech 204 00:07:06,600 --> 00:07:08,800 which brings us back to the infrastructure stack 205 00:07:09,000 --> 00:07:12,300 it seems to be shaking out into three main layers tools 206 00:07:12,300 --> 00:07:15,766 data and orchestration why those three 207 00:07:15,766 --> 00:07:17,466 why are they becoming the key pillars 208 00:07:17,466 --> 00:07:20,400 well think about what you need to build these things 209 00:07:20,633 --> 00:07:23,633 you need the tools to give the agents capabilities 210 00:07:23,633 --> 00:07:25,633 letting them browse authenticate 211 00:07:25,633 --> 00:07:27,266 use other software 212 00:07:27,266 --> 00:07:30,600 you need data for their memory and context 213 00:07:30,633 --> 00:07:32,800 and you need orchestration to manage them 214 00:07:32,800 --> 00:07:34,433 especially when you have multiple agents 215 00:07:34,433 --> 00:07:35,366 working together right 216 00:07:35,366 --> 00:07:37,100 capabilities memory management 217 00:07:37,100 --> 00:07:40,666 and the innovation within each layer is intense 218 00:07:40,666 --> 00:07:41,533 right now 219 00:07:41,666 --> 00:07:44,666 people are racing to solve the specific problems AI 220 00:07:44,666 --> 00:07:45,433 agents create 221 00:07:45,433 --> 00:07:47,600 which often aren't the same as traditional software 222 00:07:47,600 --> 00:07:48,200 problem okay 223 00:07:48,200 --> 00:07:50,333 let's break those down tools first 224 00:07:50,333 --> 00:07:52,000 you said this has seen huge growth 225 00:07:52,000 --> 00:07:53,700 yeah it feels like it's exploded 226 00:07:53,700 --> 00:07:54,166 yeah 227 00:07:54,166 --> 00:07:57,100 giving agents the ability to do things in the world 228 00:07:57,100 --> 00:08:00,600 a big piece is browser infrastructure and UI automation 229 00:08:00,900 --> 00:08:04,066 why is letting agents use websites so critical 230 00:08:04,166 --> 00:08:04,633 well 231 00:08:04,633 --> 00:08:06,633 just think about how much happens in a browser window 232 00:08:06,633 --> 00:08:07,800 right for work 233 00:08:07,800 --> 00:08:08,900 for personal stuff 234 00:08:08,900 --> 00:08:11,566 if agents can't interact with those visual interfaces 235 00:08:11,733 --> 00:08:14,400 they're locked out of a huge part of the digital world 236 00:08:14,400 --> 00:08:16,700 they can't just rely on APIs for everything 237 00:08:16,700 --> 00:08:19,366 good point APIs aren't always available or sufficient 238 00:08:19,366 --> 00:08:22,066 and so we see companies like Browser Base Light Panda 239 00:08:22,066 --> 00:08:24,600 Browserless building that core tech 240 00:08:25,000 --> 00:08:27,233 and then things like stage hand on top 241 00:08:27,233 --> 00:08:30,066 offering easier ways to code common interactions 242 00:08:30,166 --> 00:08:31,666 so you have the low level plumbing 243 00:08:31,666 --> 00:08:33,700 and the higher level components exactly 244 00:08:33,700 --> 00:08:34,900 another critical tool area 245 00:08:34,900 --> 00:08:37,066 hmm authentication and security 246 00:08:37,066 --> 00:08:38,866 yeah this feels tricky definitely 247 00:08:39,000 --> 00:08:41,133 if an agent is acting for a user 248 00:08:41,200 --> 00:08:42,933 how do permissions work 249 00:08:43,000 --> 00:08:46,066 the old username password model doesn't quite fit 250 00:08:46,100 --> 00:08:48,200 what are the big security headaches here 251 00:08:48,200 --> 00:08:49,366 well mainly 252 00:08:49,366 --> 00:08:51,200 ensuring the agent only does what the user 253 00:08:51,200 --> 00:08:52,700 actually allowed it to do 254 00:08:52,700 --> 00:08:54,766 you need fine grained control 255 00:08:55,166 --> 00:08:57,200 agent specific permission models right 256 00:08:57,200 --> 00:08:57,833 you don't want your agent 257 00:08:57,833 --> 00:08:59,300 go and row with your credentials 258 00:08:59,300 --> 00:09:00,033 exactly so 259 00:09:00,033 --> 00:09:02,166 companies like clerk and on statics 260 00:09:02,166 --> 00:09:03,966 dot AI are jumping in 261 00:09:03,966 --> 00:09:07,333 building this whole off for agents category makes sense 262 00:09:07,566 --> 00:09:09,100 okay what else in the tools layer 263 00:09:09,100 --> 00:09:10,100 tool discovery 264 00:09:10,100 --> 00:09:12,500 how does an agent even find out what tools it can use 265 00:09:12,500 --> 00:09:13,700 yeah that's huge 266 00:09:13,700 --> 00:09:15,500 right now it seems a bit ad hoc 267 00:09:15,633 --> 00:09:18,166 hmm but there's a lot of talk about Anthropics Model 268 00:09:18,166 --> 00:09:19,866 Context Protocol MCP 269 00:09:19,966 --> 00:09:22,233 people are calling it potentially the TCP I 270 00:09:22,233 --> 00:09:24,466 P for AI agents whoa 271 00:09:24,466 --> 00:09:27,000 t C P I P for agents that's a big claim 272 00:09:27,000 --> 00:09:28,400 what would a standard like m C 273 00:09:28,400 --> 00:09:32,266 P actually do the hope is it provides a common language 274 00:09:32,266 --> 00:09:34,800 a standard way for agents to discover tools 275 00:09:34,800 --> 00:09:36,200 understand how to use them 276 00:09:36,200 --> 00:09:38,733 and pass context along reliably 277 00:09:38,900 --> 00:09:40,766 and the fact that big names 278 00:09:40,766 --> 00:09:42,400 Stripe Neo 4 j 279 00:09:42,400 --> 00:09:45,466 Cloudflare are already setting up MCP servers 280 00:09:46,233 --> 00:09:48,833 that gives us some weight that adoption is key 281 00:09:48,833 --> 00:09:50,566 it could really kick start an ecosystem 282 00:09:50,566 --> 00:09:53,100 where agents and tools just work together better 283 00:09:53,100 --> 00:09:55,400 exactly but protocols can be low level 284 00:09:55,400 --> 00:09:57,866 so we're also seeing abstraction layers built on top 285 00:09:57,866 --> 00:10:00,633 like middleware sort of companies like Composio 286 00:10:00,633 --> 00:10:01,600 they offer MCP 287 00:10:01,600 --> 00:10:03,633 compatible ways to access things like Gmail 288 00:10:03,633 --> 00:10:04,433 or linear 289 00:10:04,466 --> 00:10:07,800 but through standard SDK's like typescript or Python 290 00:10:07,966 --> 00:10:09,400 what's the advantage for developers there 291 00:10:09,400 --> 00:10:11,033 oh that makes it way easier right 292 00:10:11,033 --> 00:10:13,566 you don't need to understand the nitty gritty of MCP 293 00:10:13,566 --> 00:10:16,900 itself you just use a familiar library or API call 294 00:10:16,900 --> 00:10:18,066 faster development right 295 00:10:18,066 --> 00:10:19,866 and Arcade Dev is doing something similar 296 00:10:19,866 --> 00:10:21,233 but maybe broader 297 00:10:21,233 --> 00:10:23,766 a unified API for off and tool management 298 00:10:23,766 --> 00:10:25,633 even compatible with open AI stuff 299 00:10:25,633 --> 00:10:28,266 so you get choices depending on how deep you want to go 300 00:10:28,266 --> 00:10:30,400 yeah this layered approach seems smart 301 00:10:30,466 --> 00:10:31,600 it caters to different needs 302 00:10:31,600 --> 00:10:33,966 it's definitely better than the current situation 303 00:10:34,033 --> 00:10:36,800 which feels like finding tools is mostly just asking 304 00:10:36,800 --> 00:10:38,266 around word of mouth 305 00:10:38,266 --> 00:10:39,366 uh huh yeah 306 00:10:39,366 --> 00:10:40,500 these protocols and layers 307 00:10:40,500 --> 00:10:42,200 could make it much more systematic 308 00:10:42,700 --> 00:10:45,466 okay let's shift gears to the data layer 309 00:10:45,466 --> 00:10:46,966 memory at scale 310 00:10:47,666 --> 00:10:48,700 this really shows 311 00:10:48,700 --> 00:10:51,066 how existing infrastructure has to change 312 00:10:51,266 --> 00:10:52,433 Neon's experience again 313 00:10:52,433 --> 00:10:55,000 agents creating databases 4 x faster 314 00:10:55,100 --> 00:10:56,566 that creates totally new demands 315 00:10:56,566 --> 00:10:56,900 doesn't it 316 00:10:56,900 --> 00:10:58,700 what kind of pressures does that put on their systems 317 00:10:58,700 --> 00:11:01,600 well you suddenly need instant database provisioning 318 00:11:01,600 --> 00:11:04,300 you need it to scale up massively automatically 319 00:11:04,433 --> 00:11:05,900 and you need strong isolation 320 00:11:05,900 --> 00:11:08,300 so one agent's database doesn't mess with another's 321 00:11:08,400 --> 00:11:09,833 think about that create dot X y 322 00:11:09,833 --> 00:11:10,766 z example yeah 323 00:11:10,766 --> 00:11:14,033 the developer agent spinning up 20,000 data bases 324 00:11:14,033 --> 00:11:16,733 in 36 hours just based on user prompts 325 00:11:16,866 --> 00:11:19,566 that demands incredible agility and scale 326 00:11:19,566 --> 00:11:22,166 from the data platform mind boggling scale 327 00:11:22,166 --> 00:11:23,233 and within this data layer 328 00:11:23,233 --> 00:11:24,866 we're seeing specialization too right 329 00:11:24,866 --> 00:11:26,200 like specific memory system 330 00:11:26,200 --> 00:11:29,600 exactly things like Memzero or Zep 331 00:11:29,766 --> 00:11:32,033 they're built specifically for agents to 332 00:11:32,033 --> 00:11:34,866 remember stuff over time to recall context 333 00:11:34,866 --> 00:11:36,866 how are they different from just using say 334 00:11:36,866 --> 00:11:38,800 Redis or a regular database 335 00:11:39,166 --> 00:11:39,400 well 336 00:11:39,400 --> 00:11:42,433 traditional databases are usually about structured data 337 00:11:42,433 --> 00:11:44,300 right transactions 338 00:11:44,400 --> 00:11:46,666 these agent memory systems are often geared towards 339 00:11:46,666 --> 00:11:48,733 maybe more unstructured info 340 00:11:48,800 --> 00:11:51,300 and optimized for that kind of contextual recall 341 00:11:51,300 --> 00:11:54,466 agents need for longer conversations or tasks 342 00:11:54,466 --> 00:11:56,833 got it and then there's storage itself 343 00:11:56,833 --> 00:11:59,500 we see traditional players like neon adapting 344 00:11:59,500 --> 00:12:01,266 and vector databases like pinecone 345 00:12:01,266 --> 00:12:02,800 becoming super important right 346 00:12:02,800 --> 00:12:05,066 why are vector databases such a good fit here 347 00:12:05,066 --> 00:12:07,366 because agents deal with so much unstructured stuff 348 00:12:07,366 --> 00:12:08,666 text images etc 349 00:12:08,666 --> 00:12:11,100 vector databases are great at finding similar things 350 00:12:11,100 --> 00:12:13,066 in that kind of data semantic search 351 00:12:13,066 --> 00:12:15,800 finding relevant context that's their superpower 352 00:12:15,800 --> 00:12:18,000 it's crucial for making agents smart right 353 00:12:18,000 --> 00:12:20,866 and we can't forget E P L extract transform load 354 00:12:20,866 --> 00:12:23,100 new services are popping up just to handle 355 00:12:23,100 --> 00:12:25,233 pulling in and cleaning up unstructured data 356 00:12:25,233 --> 00:12:27,066 for these agent systems makes sense 357 00:12:27,066 --> 00:12:29,766 garbage in garbage out still applies even for AI 358 00:12:29,766 --> 00:12:31,766 especially for AI okay 359 00:12:31,766 --> 00:12:33,666 final layer orchestration 360 00:12:34,100 --> 00:12:35,800 managing the complexity yeah 361 00:12:35,800 --> 00:12:37,600 once you start using multiple agents 362 00:12:37,600 --> 00:12:39,800 maybe specialized ones talking to each other 363 00:12:40,066 --> 00:12:41,433 it gets complicated fast 364 00:12:41,433 --> 00:12:43,433 you need something to manage the whole flow 365 00:12:43,433 --> 00:12:44,866 that's where tools like Lane Graph 366 00:12:44,866 --> 00:12:49,100 Crew AI let it fit in managed orchestration exactly 367 00:12:49,233 --> 00:12:51,066 they give you frameworks to design 368 00:12:51,066 --> 00:12:53,166 and run these multi agent workflows 369 00:12:53,300 --> 00:12:55,100 making sure they talk correctly 370 00:12:55,100 --> 00:12:57,500 handle steps in order recover from errors 371 00:12:57,800 --> 00:12:59,000 what are the main headaches 372 00:12:59,000 --> 00:13:01,100 when you try to coordinate multiple agents 373 00:13:01,100 --> 00:13:01,966 no lots 374 00:13:02,033 --> 00:13:04,566 making sure Agent B waits for Agent a to finish 375 00:13:04,833 --> 00:13:06,700 handling it if one agent fails 376 00:13:06,766 --> 00:13:08,766 sharing information between them reliably 377 00:13:08,766 --> 00:13:10,500 keeping track of the overall goal 378 00:13:10,666 --> 00:13:13,500 orchestration tools try to hide some of that messiness 379 00:13:13,500 --> 00:13:15,566 right and for long running processes 380 00:13:15,566 --> 00:13:17,100 you need state persistence 381 00:13:17,100 --> 00:13:19,300 which is where things like ingest hatchet 382 00:13:19,300 --> 00:13:20,400 hemorrhoid come in 383 00:13:20,400 --> 00:13:22,366 they make sure the system remembers where it was 384 00:13:22,366 --> 00:13:24,666 even if a process takes hours or days 385 00:13:24,800 --> 00:13:28,166 reliability for complex long running agent jobs 386 00:13:28,166 --> 00:13:30,833 so looking back that analogy from a year ago 387 00:13:30,833 --> 00:13:33,633 developers speeding across a half finished bridge 388 00:13:33,633 --> 00:13:34,800 yeah I like that one 389 00:13:34,800 --> 00:13:37,000 it felt accurate then and today 390 00:13:37,100 --> 00:13:38,800 the bridge is definitely more built out 391 00:13:38,800 --> 00:13:41,366 there's more structure but the traffic on the bridge 392 00:13:41,366 --> 00:13:43,866 agent adoption is growing exponentially 393 00:13:43,866 --> 00:13:46,600 good point so the construction isn't over 394 00:13:46,600 --> 00:13:49,433 it has to keep pace with the ever increasing load 395 00:13:49,433 --> 00:13:52,233 exactly which I think is the main point here 396 00:13:52,233 --> 00:13:54,466 this rapid infrastructure evolution 397 00:13:54,466 --> 00:13:56,033 it's not just incremental improvement 398 00:13:56,033 --> 00:13:57,366 it feels like a fundamental 399 00:13:57,366 --> 00:13:59,300 shift in how software gets built 400 00:13:59,366 --> 00:14:03,200 a whole new paradigm emerging and this maturing stack 401 00:14:03,200 --> 00:14:05,033 the tools data orchestration 402 00:14:05,033 --> 00:14:06,966 the protocols like some damn CP 403 00:14:06,966 --> 00:14:09,566 it's all scrambling to keep up to enable the shift 404 00:14:09,633 --> 00:14:11,166 but even with all this progress 405 00:14:11,666 --> 00:14:13,033 feels like we're still pretty early days right 406 00:14:13,033 --> 00:14:13,700 oh absolutely 407 00:14:13,700 --> 00:14:14,933 very early innings 408 00:14:14,966 --> 00:14:16,700 the potential for new ideas and innovation 409 00:14:16,700 --> 00:14:18,700 across all these layers tools 410 00:14:18,700 --> 00:14:19,966 data orchestration 411 00:14:19,966 --> 00:14:22,600 it's still massive huge opportunities remain 412 00:14:22,700 --> 00:14:24,800 okay so let's quickly recap the main points 413 00:14:24,800 --> 00:14:25,866 from our deep dive 414 00:14:25,900 --> 00:14:28,600 we saw the crazy growth of companies using agents 415 00:14:28,600 --> 00:14:31,200 the shift from just experimenting to using clear 416 00:14:31,200 --> 00:14:32,266 repeatable patterns 417 00:14:32,266 --> 00:14:34,000 and how the infrastructure is solidifying 418 00:14:34,000 --> 00:14:36,733 around those three key layers tools 419 00:14:36,800 --> 00:14:38,466 data and orchestration 420 00:14:38,466 --> 00:14:40,666 and each one is buzzing with activity 421 00:14:40,666 --> 00:14:42,000 and all this evolving 422 00:14:42,000 --> 00:14:43,833 tech is paving the way for completely 423 00:14:43,833 --> 00:14:45,200 new kinds of applications 424 00:14:45,200 --> 00:14:47,600 and services we probably haven't even thought of yet 425 00:14:47,600 --> 00:14:48,400 definitely 426 00:14:48,566 --> 00:14:51,300 so maybe a final thought to leave everyone with 427 00:14:51,466 --> 00:14:54,100 how will these big infrastructure shifts 428 00:14:54,233 --> 00:14:57,666 actually change the apps and services we use every day 429 00:14:57,966 --> 00:14:59,600 think beyond just code 430 00:14:59,966 --> 00:15:02,033 what new things become possible in 431 00:15:02,033 --> 00:15:05,066 say healthcare or education or entertainment 432 00:15:05,266 --> 00:15:07,433 now that this foundation is getting stronger 433 00:15:07,433 --> 00:15:09,666 it's a really exciting time to watch this space 434 00:15:09,666 --> 00:15:10,766 absolutely fascinating 435 00:15:10,766 --> 00:15:12,466 thanks for diving deep with us today