1 00:00:15,519 --> 00:00:17,760 Andrew. I'm the chief of software at 2 00:00:17,760 --> 00:00:19,439 Verscell 3 00:00:19,439 --> 00:00:21,679 and I'm here to talk to you about how we 4 00:00:21,679 --> 00:00:24,719 solved agent building at Verscell. I'm 5 00:00:24,719 --> 00:00:26,320 the chief of software. So I work on a 6 00:00:26,320 --> 00:00:28,320 mix of internal engineering, external 7 00:00:28,320 --> 00:00:30,399 experimentation, and generally being at 8 00:00:30,399 --> 00:00:32,320 the frontier and building new libraries, 9 00:00:32,320 --> 00:00:34,880 frameworks, and technologies. For those 10 00:00:34,880 --> 00:00:37,200 of you that don't know Verscell, 11 00:00:37,200 --> 00:00:39,520 Verscell builds a gentic infrastructure 12 00:00:39,520 --> 00:00:42,159 so people can build what's next. We get 13 00:00:42,159 --> 00:00:43,840 started in the web world helping people 14 00:00:43,840 --> 00:00:46,159 ship websites and web apps without 15 00:00:46,159 --> 00:00:47,679 having to worry about the infrastructure 16 00:00:47,679 --> 00:00:49,200 that doesn't make their app any better. 17 00:00:49,200 --> 00:00:50,799 It can scale to a million and scale down 18 00:00:50,799 --> 00:00:53,679 to zero effortlessly. But we're seeing a 19 00:00:53,679 --> 00:00:55,759 change in what people want to build. You 20 00:00:55,759 --> 00:00:57,359 know, people started by building pages, 21 00:00:57,359 --> 00:00:58,799 but now we see them want to build 22 00:00:58,799 --> 00:01:00,878 agents. And we've been embarking on a 23 00:01:00,878 --> 00:01:02,719 similar journey to make it easy for 24 00:01:02,719 --> 00:01:05,359 people to build agents and agentic 25 00:01:05,359 --> 00:01:08,239 applications easier. We built this thing 26 00:01:08,239 --> 00:01:10,079 called the AIDK. So instead of needing 27 00:01:10,079 --> 00:01:13,359 to switch out 300 400 lines of provider 28 00:01:13,359 --> 00:01:14,799 specific code, you're just going to 29 00:01:14,799 --> 00:01:16,560 switch out one line of code and we have 30 00:01:16,560 --> 00:01:18,799 the same model interface underlying for 31 00:01:18,799 --> 00:01:21,118 all these different providers. 32 00:01:21,118 --> 00:01:22,719 We built a lot of other tools to make it 33 00:01:22,719 --> 00:01:25,118 easier to have model fallbacks, secure 34 00:01:25,118 --> 00:01:27,359 code execution, better pricing when it's 35 00:01:27,359 --> 00:01:29,519 inactive and waiting for responses, as 36 00:01:29,519 --> 00:01:32,879 well as for durability and resumability. 37 00:01:32,879 --> 00:01:35,359 And I'm here to talk to you about how I 38 00:01:35,359 --> 00:01:37,118 went on this crazy experiment roughly a 39 00:01:37,118 --> 00:01:41,200 year ago that led to a aentic explosion 40 00:01:41,200 --> 00:01:43,118 at Verscell and led to a really cool 41 00:01:43,118 --> 00:01:46,159 thing that we built recently 42 00:01:46,159 --> 00:01:49,680 about uh this is 1980. Uh, Bill Gates 43 00:01:49,680 --> 00:01:52,078 before my time had this quote saying he 44 00:01:52,078 --> 00:01:53,920 imagined there would be a computer on 45 00:01:53,920 --> 00:01:55,920 every desk and in every home. You know, 46 00:01:55,920 --> 00:01:57,519 that was probably pretty contrarian then 47 00:01:57,519 --> 00:02:00,478 and today it seems like very normal to 48 00:02:00,478 --> 00:02:03,840 have that happen. And me and the CTO had 49 00:02:03,840 --> 00:02:06,078 this thought, you know, instead of a 50 00:02:06,078 --> 00:02:08,080 computer on every desk, could we 51 00:02:08,080 --> 00:02:10,878 potentially have an agent on every desk? 52 00:02:10,878 --> 00:02:13,199 You know, today we only really use 53 00:02:13,199 --> 00:02:15,520 agents for coding and technical 54 00:02:15,520 --> 00:02:17,120 workloads, but we're starting to see 55 00:02:17,120 --> 00:02:19,120 expansion into things like design, 56 00:02:19,120 --> 00:02:22,400 product management, and other verticals. 57 00:02:22,400 --> 00:02:24,000 And this was maybe about a year ago, so 58 00:02:24,000 --> 00:02:25,598 I would say I'm pretty early to this, 59 00:02:25,598 --> 00:02:27,520 but that was when it was like sonnet 4 60 00:02:27,520 --> 00:02:29,120 and things weren't as sophisticated as 61 00:02:29,120 --> 00:02:32,560 they were today. And I tried to actually 62 00:02:32,560 --> 00:02:34,080 explore this out, see what we could do 63 00:02:34,080 --> 00:02:36,239 about it. I went around to various job 64 00:02:36,239 --> 00:02:39,120 functions at Verscell, marketing, sales, 65 00:02:39,120 --> 00:02:41,680 finance, legal, and I asked them, what 66 00:02:41,680 --> 00:02:45,120 do you hate most about your job? And the 67 00:02:45,120 --> 00:02:47,280 most compelling use case I heard was 68 00:02:47,280 --> 00:02:50,080 that the data team, they were growing. 69 00:02:50,080 --> 00:02:52,400 They were a very lean team, but Versel 70 00:02:52,400 --> 00:02:54,479 was growing faster. You know, they had 71 00:02:54,479 --> 00:02:56,800 so much more data from customers, 72 00:02:56,800 --> 00:02:59,840 analytics, metrics, sales. They just had 73 00:02:59,840 --> 00:03:01,598 to keep on aggregating and keep on 74 00:03:01,598 --> 00:03:05,199 making available for themselves to use. 75 00:03:05,199 --> 00:03:07,039 And at this time, if you think about 76 00:03:07,039 --> 00:03:08,800 what the data science people ever have 77 00:03:08,800 --> 00:03:10,878 to do, whenever someone from marketing 78 00:03:10,878 --> 00:03:12,959 or sales has a question about a customer 79 00:03:12,959 --> 00:03:15,360 or product, the data science team has to 80 00:03:15,360 --> 00:03:17,280 drop everything they're doing, write the 81 00:03:17,280 --> 00:03:19,519 query, process it, do an analysis, and 82 00:03:19,519 --> 00:03:21,039 come back with some recommendation on 83 00:03:21,039 --> 00:03:24,080 what to do. And this was really killer 84 00:03:24,080 --> 00:03:26,000 to productivity. You know, the data team 85 00:03:26,000 --> 00:03:28,000 did not want to drop everything and just 86 00:03:28,000 --> 00:03:29,919 write queries all day. And so I worked 87 00:03:29,919 --> 00:03:31,919 with our VP of data to try to build a 88 00:03:31,919 --> 00:03:36,000 better way for them to operate this way. 89 00:03:36,000 --> 00:03:37,919 And so if you think about the very first 90 00:03:37,919 --> 00:03:39,519 thing you would ever do if you want to 91 00:03:39,519 --> 00:03:41,519 try to use AI to solve a problem, you 92 00:03:41,519 --> 00:03:43,840 may just build like a huge mega prompt. 93 00:03:43,840 --> 00:03:46,158 You know, you just have a question, you 94 00:03:46,158 --> 00:03:48,318 pass into an LM, you have it respond, 95 00:03:48,318 --> 00:03:50,158 and that's it. You know, this was how 96 00:03:50,158 --> 00:03:51,598 the first version really looked. 97 00:03:51,598 --> 00:03:54,400 Honestly, I asked them for a dump of of 98 00:03:54,400 --> 00:03:57,360 the snowflake schema. I pasted it into a 99 00:03:57,360 --> 00:03:59,280 system prompt with a question and then 100 00:03:59,280 --> 00:04:01,199 when it generated SQL I actually copy 101 00:04:01,199 --> 00:04:02,799 and pasted that in and just ran it 102 00:04:02,799 --> 00:04:04,318 myself. You know I just want to see are 103 00:04:04,318 --> 00:04:06,799 the models good enough today in order to 104 00:04:06,799 --> 00:04:09,519 write valid SQL given some decent 105 00:04:09,519 --> 00:04:11,759 structure. And I would say this gave us 106 00:04:11,759 --> 00:04:13,280 a little bit of confidence that you know 107 00:04:13,280 --> 00:04:15,120 models today aren't that good but maybe 108 00:04:15,120 --> 00:04:17,439 we can harness engineer or make the 109 00:04:17,439 --> 00:04:19,519 context around it a little better and 110 00:04:19,519 --> 00:04:21,120 give us some more guardrails to operate 111 00:04:21,120 --> 00:04:22,720 a little better. 112 00:04:22,720 --> 00:04:24,000 And so if you actually think about what 113 00:04:24,000 --> 00:04:25,519 a data scientist actually needs to do 114 00:04:25,519 --> 00:04:27,279 when they get a question, you know, they 115 00:04:27,279 --> 00:04:29,360 have to process the question, they may 116 00:04:29,360 --> 00:04:31,120 have to explore the semantic layer and 117 00:04:31,120 --> 00:04:32,319 actually figure out what the join 118 00:04:32,319 --> 00:04:34,720 patterns are. They will actually go and 119 00:04:34,720 --> 00:04:37,120 execute the SQL. They may go back and do 120 00:04:37,120 --> 00:04:39,439 that again if the SQL did not execute or 121 00:04:39,439 --> 00:04:40,959 was too expensive. And they'll 122 00:04:40,959 --> 00:04:42,399 eventually report on it, including 123 00:04:42,399 --> 00:04:44,319 visualize the data, maybe write some 124 00:04:44,319 --> 00:04:46,240 paragraphs, maybe do a retro, maybe do 125 00:04:46,240 --> 00:04:48,079 some other stuff. 126 00:04:48,079 --> 00:04:49,199 And so if you think about those 127 00:04:49,199 --> 00:04:51,040 different phases, me and the VP of data 128 00:04:51,040 --> 00:04:53,279 tried to sit down and map those out into 129 00:04:53,279 --> 00:04:55,680 specific agent workloads. And so the 130 00:04:55,680 --> 00:04:57,680 second version of this data science 131 00:04:57,680 --> 00:04:59,839 agent uh called D0. I'm going to 132 00:04:59,839 --> 00:05:02,478 reference D0 from now on is you ask a 133 00:05:02,478 --> 00:05:04,478 question. We have a query agent that 134 00:05:04,478 --> 00:05:06,800 passes on a query to the planning agent 135 00:05:06,800 --> 00:05:08,800 that will then have an execution agent 136 00:05:08,800 --> 00:05:11,120 etc. And if you chain all of these 137 00:05:11,120 --> 00:05:12,560 together, you actually get something 138 00:05:12,560 --> 00:05:14,639 that looks like this where each agent 139 00:05:14,639 --> 00:05:16,560 has a very dedicated system prompt 140 00:05:16,560 --> 00:05:18,478 focused to what that does with tools 141 00:05:18,478 --> 00:05:20,879 scoped to exactly that function. So 142 00:05:20,879 --> 00:05:23,038 example here, you can see that for the 143 00:05:23,038 --> 00:05:25,839 first one, the planning agent has a read 144 00:05:25,839 --> 00:05:29,199 entity YAML and a and a search schemas 145 00:05:29,199 --> 00:05:32,079 tool. And so it will only use those 146 00:05:32,079 --> 00:05:34,000 capabilities until it has an answer to 147 00:05:34,000 --> 00:05:36,160 pass on to the planning agent and then 148 00:05:36,160 --> 00:05:38,800 to the SQL agent and then to reporting. 149 00:05:38,800 --> 00:05:40,560 And this was getting better. You know, 150 00:05:40,560 --> 00:05:42,959 we were able to get away from having to 151 00:05:42,959 --> 00:05:45,038 copy and paste a SQL and have to come 152 00:05:45,038 --> 00:05:46,720 back and report on it. It was now 153 00:05:46,720 --> 00:05:48,399 actually doing like the end to end loop 154 00:05:48,399 --> 00:05:51,600 from question to answer. 155 00:05:51,600 --> 00:05:54,639 But we started hitting some walls with 156 00:05:54,639 --> 00:05:56,639 this architecture. And around this time 157 00:05:56,639 --> 00:05:58,478 we came to the conclusion that you know 158 00:05:58,478 --> 00:05:59,918 what you actually need is you need one 159 00:05:59,918 --> 00:06:01,839 agent with all the mega context within 160 00:06:01,839 --> 00:06:04,319 it and for it to sort of manage its own 161 00:06:04,319 --> 00:06:06,399 memory. You know this was around the 162 00:06:06,399 --> 00:06:10,000 time when we realized that you want to 163 00:06:10,000 --> 00:06:11,279 actually have the agent be able to look 164 00:06:11,279 --> 00:06:13,120 back on what it's done sort of reflect 165 00:06:13,120 --> 00:06:15,199 and figure out the steps that got to get 166 00:06:15,199 --> 00:06:17,038 here. And with the previous model you 167 00:06:17,038 --> 00:06:18,879 may have noticed that the only thing 168 00:06:18,879 --> 00:06:21,279 that the next agent gets is a summary 169 00:06:21,279 --> 00:06:22,879 and a small snippet of the previous 170 00:06:22,879 --> 00:06:24,959 thing that was done. Now this way you 171 00:06:24,959 --> 00:06:27,360 can imagine that you have one mega agent 172 00:06:27,360 --> 00:06:29,600 and internally it manages its own state. 173 00:06:29,600 --> 00:06:31,038 At some points it's planning, some 174 00:06:31,038 --> 00:06:32,399 points it's building, some points it's 175 00:06:32,399 --> 00:06:33,839 executing and some points it's 176 00:06:33,839 --> 00:06:36,399 reporting. And this is sort of what it 177 00:06:36,399 --> 00:06:38,639 looked like. You know you have one big 178 00:06:38,639 --> 00:06:42,240 AI call maybe max steps 100 and you give 179 00:06:42,240 --> 00:06:44,240 it the ability to manage its own state 180 00:06:44,240 --> 00:06:46,800 based on where it's at inside of its 181 00:06:46,800 --> 00:06:48,879 execution journey. And so you can see 182 00:06:48,879 --> 00:06:50,160 similar tools, you can see a similar 183 00:06:50,160 --> 00:06:51,918 shape, but the best part about this is 184 00:06:51,918 --> 00:06:53,600 if it ever ran to an error when 185 00:06:53,600 --> 00:06:55,600 executing or joining, it could go back 186 00:06:55,600 --> 00:06:57,680 and explore more or it could go and read 187 00:06:57,680 --> 00:06:59,120 more and figure out what it was doing 188 00:06:59,120 --> 00:07:01,839 wrong. And it was very good at this 189 00:07:01,839 --> 00:07:04,478 point. We were pretty confident in the 190 00:07:04,478 --> 00:07:06,399 actual system at hand and we actually 191 00:07:06,399 --> 00:07:08,800 spread it to a few trusted members ever. 192 00:07:08,800 --> 00:07:10,079 You know, this is a very powerful tool 193 00:07:10,079 --> 00:07:11,439 and we didn't really want to put in the 194 00:07:11,439 --> 00:07:12,879 hands of the wrong people or people that 195 00:07:12,879 --> 00:07:15,279 were using very critical workloads. So, 196 00:07:15,279 --> 00:07:17,120 we got to a few people's hands and the 197 00:07:17,120 --> 00:07:18,879 immediate response was it was awful. You 198 00:07:18,879 --> 00:07:20,240 know, we thought we were cooking. We 199 00:07:20,240 --> 00:07:22,240 thought this was, you know, nailing 30% 200 00:07:22,240 --> 00:07:24,160 of our evals, but we couldn't have 201 00:07:24,160 --> 00:07:25,839 anticipated some of the questions that 202 00:07:25,839 --> 00:07:27,839 were being asked. And for us to spend 203 00:07:27,839 --> 00:07:29,839 more time manually mapping out some of 204 00:07:29,839 --> 00:07:32,399 these scenarios, it didn't seem like a 205 00:07:32,399 --> 00:07:35,199 very scalable way to do this. 206 00:07:35,199 --> 00:07:39,120 And then claude code and opus 4.5 came 207 00:07:39,120 --> 00:07:41,680 out. Well, more like Opus 4.5 came out 208 00:07:41,680 --> 00:07:43,759 and it in tangent with claw code which 209 00:07:43,759 --> 00:07:46,240 is so powerful. You know they sort of 210 00:07:46,240 --> 00:07:48,959 unlocked the concept of a file system 211 00:07:48,959 --> 00:07:52,560 agent and we on the side were like wow 212 00:07:52,560 --> 00:07:55,199 clawed code and Opus 4.5 is basically 213 00:07:55,199 --> 00:07:57,439 AGI compared to what we had before. You 214 00:07:57,439 --> 00:07:58,879 know it would answer most of our 215 00:07:58,879 --> 00:08:02,000 questions without even without even 216 00:08:02,000 --> 00:08:03,918 missing a beat um compared to the 217 00:08:03,918 --> 00:08:05,918 handgrown agent we had. And when we 218 00:08:05,918 --> 00:08:07,439 tried to step back and wonder what we 219 00:08:07,439 --> 00:08:08,879 were doing wrong and why this was so 220 00:08:08,879 --> 00:08:11,598 much better, we realized that the big 221 00:08:11,598 --> 00:08:13,759 unlock was that it was just a file 222 00:08:13,759 --> 00:08:15,759 system. You know, we it had a very 223 00:08:15,759 --> 00:08:17,918 minimal set of tools, list file, read 224 00:08:17,918 --> 00:08:20,959 file, run bash, and we gave a few more 225 00:08:20,959 --> 00:08:24,478 here for uh our own data agent use case. 226 00:08:24,478 --> 00:08:26,478 But the biggest thing was it was able to 227 00:08:26,478 --> 00:08:28,079 use the tools that agents are well 228 00:08:28,079 --> 00:08:30,478 trained on and was able to explore and 229 00:08:30,478 --> 00:08:32,879 write work where it needs to. you know, 230 00:08:32,879 --> 00:08:35,278 we weren't giving it claw code was not 231 00:08:35,278 --> 00:08:36,719 giving it a very prescriptive set of 232 00:08:36,719 --> 00:08:38,399 tools. It was sort of just letting it go 233 00:08:38,399 --> 00:08:41,440 wild and explore emergent behavior. And 234 00:08:41,440 --> 00:08:43,278 so from this, we learned that you can 235 00:08:43,278 --> 00:08:45,200 really just use a file system. You know, 236 00:08:45,200 --> 00:08:47,200 we we saw the learnings from claw code 237 00:08:47,200 --> 00:08:49,440 and how powerful it was given that it 238 00:08:49,440 --> 00:08:52,320 just executes locally. And we tried to 239 00:08:52,320 --> 00:08:54,080 rebuild it in a way that was very cloud 240 00:08:54,080 --> 00:08:56,399 codeesque. You know, it was now going to 241 00:08:56,399 --> 00:08:58,480 run in a sandbox. That sandbox would 242 00:08:58,480 --> 00:09:00,639 dump the whole semantic layer into it. 243 00:09:00,639 --> 00:09:02,320 You could the agent would be able to 244 00:09:02,320 --> 00:09:04,480 grab, bash, read file, write file all 245 00:09:04,480 --> 00:09:06,240 around to figure out what it needs and 246 00:09:06,240 --> 00:09:07,600 we would just sprinkle a few tools on 247 00:09:07,600 --> 00:09:08,958 top to make sure it could do everything 248 00:09:08,958 --> 00:09:12,399 that is versel specific. 249 00:09:12,399 --> 00:09:14,480 And this was actually the biggest unlock 250 00:09:14,480 --> 00:09:17,360 ever. You know, the leap from single 251 00:09:17,360 --> 00:09:19,759 agent to cloud code SDK and then from 252 00:09:19,759 --> 00:09:22,559 cloud code SDK to file system agent in 253 00:09:22,559 --> 00:09:25,200 general, fine-tuned or purpose-built for 254 00:09:25,200 --> 00:09:27,839 our use case was an amazing leap. At 255 00:09:27,839 --> 00:09:28,799 this point, we were starting to get 256 00:09:28,799 --> 00:09:31,039 ready to give it away to more people at 257 00:09:31,039 --> 00:09:32,639 Versell. 258 00:09:32,639 --> 00:09:35,600 And at this point, the eval score 259 00:09:35,600 --> 00:09:37,759 basically doubled. And I wrote this uh 260 00:09:37,759 --> 00:09:39,519 this is basically how it looks. Um it's 261 00:09:39,519 --> 00:09:41,039 very simple. You just give it a bash 262 00:09:41,039 --> 00:09:42,879 tool. We have a nice helper called bash 263 00:09:42,879 --> 00:09:45,120 tool on npm and you attach it to a 264 00:09:45,120 --> 00:09:46,958 sandbox and you can attach files to the 265 00:09:46,958 --> 00:09:48,559 sandbox for it to read, write and 266 00:09:48,559 --> 00:09:50,480 execute. 267 00:09:50,480 --> 00:09:52,320 And after this revelation and after I 268 00:09:52,320 --> 00:09:54,639 saw that we were passing so many of the 269 00:09:54,639 --> 00:09:56,480 questions that we failed to do before, I 270 00:09:56,480 --> 00:09:58,879 wrote this banger blog post. It's uh 271 00:09:58,879 --> 00:10:01,039 it's actually up today. And the week 272 00:10:01,039 --> 00:10:02,799 that I wrote this, it was responsible 273 00:10:02,799 --> 00:10:05,839 for 70% of our versel.com traffic. So 274 00:10:05,839 --> 00:10:09,120 you know it's a banger. And after that, 275 00:10:09,120 --> 00:10:12,399 the next logical step was that we want 276 00:10:12,399 --> 00:10:14,480 to figure out the common use cases we 277 00:10:14,480 --> 00:10:16,399 had. So by then we've already sort of 278 00:10:16,399 --> 00:10:19,278 let a leash on all of our cell and we 279 00:10:19,278 --> 00:10:21,278 were getting thousands of queries a day 280 00:10:21,278 --> 00:10:23,120 from people wanting everything from 281 00:10:23,120 --> 00:10:25,440 customer metrics sales metrics number 282 00:10:25,440 --> 00:10:28,879 metrics npm downloads and it turns out 283 00:10:28,879 --> 00:10:30,078 that a lot of these queries are actually 284 00:10:30,078 --> 00:10:31,679 the same in shape you know there's only 285 00:10:31,679 --> 00:10:33,600 so many ways you can do an aggregation 286 00:10:33,600 --> 00:10:34,958 only so many ways you can look up a 287 00:10:34,958 --> 00:10:36,240 product only so many ways you can do 288 00:10:36,240 --> 00:10:38,639 billing info and so we actually have a 289 00:10:38,639 --> 00:10:40,480 recurring job that takes the most recent 290 00:10:40,480 --> 00:10:42,480 queries and tries to distill them into a 291 00:10:42,480 --> 00:10:44,879 skill and right now we have roughly 100 292 00:10:44,879 --> 00:10:47,440 skills that do a mix of aggregation all 293 00:10:47,440 --> 00:10:49,120 the way through looking up specific data 294 00:10:49,120 --> 00:10:51,839 about certain people. And we found this 295 00:10:51,839 --> 00:10:53,120 very effective because if you think 296 00:10:53,120 --> 00:10:55,600 about every new agent run, it sort of 297 00:10:55,600 --> 00:10:56,879 just starts from nothing. You know, 298 00:10:56,879 --> 00:10:58,159 there's really no pre-established 299 00:10:58,159 --> 00:11:00,399 context besides, you know, the semantic 300 00:11:00,399 --> 00:11:02,320 layer and the system prompt. But with a 301 00:11:02,320 --> 00:11:04,159 skill, it already starts off with a lot 302 00:11:04,159 --> 00:11:05,919 of contextual knowledge that has 303 00:11:05,919 --> 00:11:09,200 otherwise already been done. 304 00:11:09,200 --> 00:11:11,120 And this is roughly how it looks. It's 305 00:11:11,120 --> 00:11:12,559 very similar to the previous one, but 306 00:11:12,559 --> 00:11:14,078 the inclusion of a skills folder is 307 00:11:14,078 --> 00:11:16,159 actually very powerful. Um, we also 308 00:11:16,159 --> 00:11:17,440 built this tool at Verscell called 309 00:11:17,440 --> 00:11:19,440 Skillsh. It's the most popular way to 310 00:11:19,440 --> 00:11:23,360 find agent skills and run them yourself. 311 00:11:23,360 --> 00:11:25,600 And I I'm saying all this because this 312 00:11:25,600 --> 00:11:27,440 journey is something that most of you 313 00:11:27,440 --> 00:11:29,200 may hit once in a while where you start 314 00:11:29,200 --> 00:11:30,639 from something simple and you gradually 315 00:11:30,639 --> 00:11:32,480 add complexity and you eventually hit a 316 00:11:32,480 --> 00:11:35,278 system in which you can ship to prod. 317 00:11:35,278 --> 00:11:37,679 And I'm telling you this because at 318 00:11:37,679 --> 00:11:39,919 every step along building this agent, 319 00:11:39,919 --> 00:11:42,639 someone Everell was agent curious and 320 00:11:42,639 --> 00:11:45,519 they tried to fork off of my DZero agent 321 00:11:45,519 --> 00:11:47,679 and build their own. And at every step, 322 00:11:47,679 --> 00:11:49,600 we sort of had a better way to do 323 00:11:49,600 --> 00:11:51,759 something that was not previously known. 324 00:11:51,759 --> 00:11:53,360 And we were wondering like what if 325 00:11:53,360 --> 00:11:55,600 people today could start from the very 326 00:11:55,600 --> 00:11:57,839 last insight and not have to ever start 327 00:11:57,839 --> 00:12:00,078 from just a simple prompt or from 328 00:12:00,078 --> 00:12:03,120 reinventing best principles from first 329 00:12:03,120 --> 00:12:04,639 principles. 330 00:12:04,639 --> 00:12:06,639 And so we actually thought what if we 331 00:12:06,639 --> 00:12:09,600 built the Nex.js for agents. For those 332 00:12:09,600 --> 00:12:11,839 that don't know, Nex.js is a popular web 333 00:12:11,839 --> 00:12:13,759 framework that Verscell built that 334 00:12:13,759 --> 00:12:16,639 invented this thing of file system uh 335 00:12:16,639 --> 00:12:18,480 framework defined infrastructure. You 336 00:12:18,480 --> 00:12:19,919 don't have to worry about where things 337 00:12:19,919 --> 00:12:22,000 go. You just have to write files in the 338 00:12:22,000 --> 00:12:24,399 right conventions and it automatically 339 00:12:24,399 --> 00:12:26,159 declares where they should go. Your 340 00:12:26,159 --> 00:12:28,320 pages go to the CDN. Your serverless 341 00:12:28,320 --> 00:12:30,399 functions go there. Your caching goes in 342 00:12:30,399 --> 00:12:32,799 the middle. And we thought, you know, 343 00:12:32,799 --> 00:12:34,559 building agents should be this simple. 344 00:12:34,559 --> 00:12:35,919 You should only have to create a skills 345 00:12:35,919 --> 00:12:37,839 folder, a tools folder, a channels 346 00:12:37,839 --> 00:12:39,120 folder, and you should be able to just 347 00:12:39,120 --> 00:12:40,879 declare these very easily. And the 348 00:12:40,879 --> 00:12:42,399 framework should know exactly how to 349 00:12:42,399 --> 00:12:45,039 make an agent out of it. 350 00:12:45,039 --> 00:12:46,879 And that's why two weeks ago we released 351 00:12:46,879 --> 00:12:49,600 Eve. Eve is a agent framework like the 352 00:12:49,600 --> 00:12:51,519 next.js GS for agents where it's very 353 00:12:51,519 --> 00:12:53,600 easy from just starting with a sample 354 00:12:53,600 --> 00:12:56,078 template to having a fully agent ready 355 00:12:56,078 --> 00:12:57,839 and being able to add in your own custom 356 00:12:57,839 --> 00:12:59,759 knowledge, your own custom tools and 357 00:12:59,759 --> 00:13:01,440 even integrated into the channels that 358 00:13:01,440 --> 00:13:04,159 you are familiar with. 359 00:13:04,159 --> 00:13:05,919 This is roughly what we think an agent 360 00:13:05,919 --> 00:13:07,440 actually looks like. You know, an agent 361 00:13:07,440 --> 00:13:09,519 has a runtime and it has channels. And 362 00:13:09,519 --> 00:13:10,879 in that runtime, you're going to have 363 00:13:10,879 --> 00:13:12,639 durability. You're going to want to run 364 00:13:12,639 --> 00:13:14,159 things in an isolate environment. You're 365 00:13:14,159 --> 00:13:15,200 going to want to call into different 366 00:13:15,200 --> 00:13:16,720 models. And you're going to want to have 367 00:13:16,720 --> 00:13:19,200 connections. And we built this with open 368 00:13:19,200 --> 00:13:21,039 source in mind. You know, we built Eve 369 00:13:21,039 --> 00:13:22,720 so you can plug in your own open source 370 00:13:22,720 --> 00:13:25,200 adapters for Postgress, OpenAI's uh 371 00:13:25,200 --> 00:13:28,639 responses API, Docker, other connectors. 372 00:13:28,639 --> 00:13:30,240 But we also made it incredibly easy to 373 00:13:30,240 --> 00:13:32,000 deploy in Verscell. The only thing here 374 00:13:32,000 --> 00:13:33,519 you see different is that everything 375 00:13:33,519 --> 00:13:35,200 here is using a Verscell product that 376 00:13:35,200 --> 00:13:36,879 we've been building over the years in 377 00:13:36,879 --> 00:13:38,399 order to make it easy to build these 378 00:13:38,399 --> 00:13:40,399 experiences. Versell workflows for 379 00:13:40,399 --> 00:13:42,320 durability, sandbox for secure 380 00:13:42,320 --> 00:13:44,000 execution, and Verscell connect, 381 00:13:44,000 --> 00:13:45,679 something we just released to make it 382 00:13:45,679 --> 00:13:48,240 easy to generate short-lived ODC tokens 383 00:13:48,240 --> 00:13:51,039 for connections. 384 00:13:51,039 --> 00:13:53,278 And we actually rewrote the whole D0ero 385 00:13:53,278 --> 00:13:56,000 agent in Eve as we were building Eve and 386 00:13:56,000 --> 00:13:57,519 from the convoluted structures behind 387 00:13:57,519 --> 00:13:59,120 the scenes that you did not see from the 388 00:13:59,120 --> 00:14:01,039 code. Um, this is roughly how the file 389 00:14:01,039 --> 00:14:02,879 system looks. It's very simple. You have 390 00:14:02,879 --> 00:14:05,039 a bunch of system instructions, a couple 391 00:14:05,039 --> 00:14:07,759 skills, a couple tools, and it's very 392 00:14:07,759 --> 00:14:09,600 easy to compose this into a real agent, 393 00:14:09,600 --> 00:14:12,320 and it's very easy to iterate on. We 394 00:14:12,320 --> 00:14:13,679 actually gave this out to a few beta 395 00:14:13,679 --> 00:14:15,278 customers before we actually fully 396 00:14:15,278 --> 00:14:16,958 released it two weeks ago at our London 397 00:14:16,958 --> 00:14:18,879 event. And this one company that 398 00:14:18,879 --> 00:14:20,799 partners closely with us, Aura. They've 399 00:14:20,799 --> 00:14:22,320 rebuilt their agent that's sort of like 400 00:14:22,320 --> 00:14:25,360 a mini claw to go and test people's 401 00:14:25,360 --> 00:14:27,839 services. It goes to websites, installs 402 00:14:27,839 --> 00:14:30,799 them, it tries to use them. And they've 403 00:14:30,799 --> 00:14:33,198 seen incredible success on building 404 00:14:33,198 --> 00:14:35,039 their own agent from the ground up using 405 00:14:35,039 --> 00:14:37,278 Eve compared to using an off-the-shelf 406 00:14:37,278 --> 00:14:40,320 cloud code. Fewer steps, better 407 00:14:40,320 --> 00:14:45,039 successes, as well as better insights. 408 00:14:45,039 --> 00:14:47,360 And when you deploy Eve to Verscell, you 409 00:14:47,360 --> 00:14:49,278 get observability observability out of 410 00:14:49,278 --> 00:14:50,799 the box. You can see here that you get 411 00:14:50,799 --> 00:14:52,399 all the agent runs, you see all the tool 412 00:14:52,399 --> 00:14:55,120 calls, you see each step it takes as 413 00:14:55,120 --> 00:14:56,799 well as maybe some estimated costs and 414 00:14:56,799 --> 00:14:58,320 some optimizations you could potentially 415 00:14:58,320 --> 00:15:00,399 take. 416 00:15:00,399 --> 00:15:03,039 And you can get start today at eve.dev. 417 00:15:03,039 --> 00:15:04,399 You can just clone it and you can just 418 00:15:04,399 --> 00:15:06,159 start a template, deploy easily, 419 00:15:06,159 --> 00:15:08,240 self-host if you need. And the reason 420 00:15:08,240 --> 00:15:10,799 why I bring this up is because I hope 421 00:15:10,799 --> 00:15:12,399 that there will be more and more 422 00:15:12,399 --> 00:15:15,198 business specific use case agents. You 423 00:15:15,198 --> 00:15:17,198 know, before we built Ezero, we actually 424 00:15:17,198 --> 00:15:20,320 battle tested a lot of the industry 425 00:15:20,320 --> 00:15:22,480 well-funded startups that were doing 426 00:15:22,480 --> 00:15:24,159 these vertical agents that were 427 00:15:24,159 --> 00:15:26,159 dedicated to taking your Snowflake 428 00:15:26,159 --> 00:15:28,639 instance and making it so their agent 429 00:15:28,639 --> 00:15:31,440 could run Snowflake queries against it. 430 00:15:31,440 --> 00:15:33,039 But we found out that what really makes 431 00:15:33,039 --> 00:15:36,480 this agent good is it has a lot of very 432 00:15:36,480 --> 00:15:39,360 specific uh company knowledge. You know, 433 00:15:39,360 --> 00:15:41,519 the way that Versel is a web- based 434 00:15:41,519 --> 00:15:43,278 company. We have a lot of customers that 435 00:15:43,278 --> 00:15:45,759 have websites and web properties. That 436 00:15:45,759 --> 00:15:48,799 goes a lot deeper into when you should 437 00:15:48,799 --> 00:15:50,639 query for what and what things link to 438 00:15:50,639 --> 00:15:52,799 what. And so a lot of these 439 00:15:52,799 --> 00:15:54,320 off-the-shelf agents, they're great. 440 00:15:54,320 --> 00:15:56,159 They're good to try, but I think if you 441 00:15:56,159 --> 00:15:57,600 really want to get the most juice out of 442 00:15:57,600 --> 00:15:58,879 a squeeze, you should really try to 443 00:15:58,879 --> 00:16:00,799 build your own agent and add in as much 444 00:16:00,799 --> 00:16:03,759 company specific knowledge as you can. 445 00:16:03,759 --> 00:16:06,879 Today, you know, we've had 20 roughly 446 00:16:06,879 --> 00:16:09,919 decently PMF agents adversel that range 447 00:16:09,919 --> 00:16:13,120 from anything from marketing retros to 448 00:16:13,120 --> 00:16:15,519 figure out who to reach out to to the 449 00:16:15,519 --> 00:16:18,399 first ever red line of a contract when 450 00:16:18,399 --> 00:16:20,799 legal sees a new negotiation all the way 451 00:16:20,799 --> 00:16:22,879 to my data science agent helping with 452 00:16:22,879 --> 00:16:26,000 with data queries. And that goes to show 453 00:16:26,000 --> 00:16:27,839 that we ever have been very 454 00:16:27,839 --> 00:16:30,320 agent-filled. You know, all of this 455 00:16:30,320 --> 00:16:32,559 stuff is actually saving us a lot of 456 00:16:32,559 --> 00:16:34,639 time. The data team has never been more 457 00:16:34,639 --> 00:16:36,639 productive. They have more time to go 458 00:16:36,639 --> 00:16:38,240 and improve the performance of 459 00:16:38,240 --> 00:16:40,159 Snowflake, to add new data sources that 460 00:16:40,159 --> 00:16:42,159 were missing, to fill in the gaps that 461 00:16:42,159 --> 00:16:43,679 they previously did not have time to 462 00:16:43,679 --> 00:16:45,278 because they were so busy writing 463 00:16:45,278 --> 00:16:48,000 queries. And I think it's never been 464 00:16:48,000 --> 00:16:50,799 easier for you at your big, small, 465 00:16:50,799 --> 00:16:53,278 medium-sized company to sort of automate 466 00:16:53,278 --> 00:16:55,278 away some of the things that you do not 467 00:16:55,278 --> 00:16:56,559 want to do or some of the things that 468 00:16:56,559 --> 00:16:58,480 you're spending too much time doing. You 469 00:16:58,480 --> 00:17:00,879 know, I think a lot of HR, finance, 470 00:17:00,879 --> 00:17:04,078 sales can be somewhat automated with 471 00:17:04,078 --> 00:17:06,240 agents. And I think Eve is the best way 472 00:17:06,240 --> 00:17:09,679 to build said agents today. 473 00:17:09,679 --> 00:17:11,679 And these are my socials. Thank you all 474 00:17:11,679 --> 00:17:13,679 for coming and listening. I'm Andrew and 475 00:17:13,679 --> 00:17:15,359 I'll be around if you want to chat 476 00:17:15,359 --> 00:17:18,359 outside.