1 00:00:00,240 --> 00:00:03,432 What It Takes to Onboard AI Agents by Anna Pinole at 2 00:00:03,456 --> 00:00:07,944 NFX as voiced by the AOK Voicebot About 3 00:00:08,032 --> 00:00:11,944 18 months. Ago, we started getting our first AI agent 4 00:00:11,992 --> 00:00:15,240 pitches. It was clear this had huge potential, 5 00:00:15,400 --> 00:00:18,696 but now we're seeing the full map in even more clarity. 6 00:00:18,888 --> 00:00:22,648 Quicker recap we see AI agents turning labor into 7 00:00:22,704 --> 00:00:25,740 software, a market size in the trillions. 8 00:00:25,880 --> 00:00:29,476 Since our first essay on this, we've worked with amazing companies 9 00:00:29,548 --> 00:00:32,756 in this space and want to do more of it. But if you're 10 00:00:32,788 --> 00:00:36,052 following this space as closely as we are, you probably 11 00:00:36,116 --> 00:00:39,172 have noticed something Progress and adoption are out 12 00:00:39,196 --> 00:00:43,508 of sync. On one hand, there is rapid technological progress. 13 00:00:43,684 --> 00:00:48,484 Just recently, Tool use operator Gemini 14 00:00:48,532 --> 00:00:54,168 2.0 and improved reasoning O3R1.3.7 15 00:00:54,224 --> 00:00:57,688 Sonnet emerged as new AI capabilities, 16 00:00:57,864 --> 00:01:02,008 both of which represent fundamental prerequisites for AI agents 17 00:01:02,104 --> 00:01:06,184 and get us closer to the future. A world where AI 18 00:01:06,232 --> 00:01:10,104 agents can act autonomously and execute complex tasks 19 00:01:10,232 --> 00:01:13,640 at a far cheaper price than we thought possible even a few months 20 00:01:13,680 --> 00:01:17,114 ago is very real. Novel capabilities 21 00:01:17,312 --> 00:01:21,126 paired with the continuous improvements in AI performance and 22 00:01:21,198 --> 00:01:25,190 cost see Deep SEQ and this are setting the foundation 23 00:01:25,270 --> 00:01:28,806 for future exploding demand. That's the good news. 24 00:01:28,958 --> 00:01:33,254 The less good news is there's still a disconnect between progress and adoption, 25 00:01:33,382 --> 00:01:37,606 a gap between the intent to implement AI at work and 26 00:01:37,678 --> 00:01:41,494 actually doing it. For example, a recent McKinsey 27 00:01:41,542 --> 00:01:44,930 survey of 100 organizations doing greater than 28 00:01:44,970 --> 00:01:48,658 $50 million in annual revenue recently found that 29 00:01:49,158 --> 00:01:53,490 63% of leaders thought implementing AI was a high priority, 30 00:01:53,650 --> 00:01:56,738 but 91% of those respondents didn't feel 31 00:01:56,794 --> 00:01:59,986 prepared to do so. It's very early days, 32 00:02:00,138 --> 00:02:03,362 and that's where you come in. Your primary job is 33 00:02:03,386 --> 00:02:06,658 to be a bridge between deep technical progress and mass 34 00:02:06,714 --> 00:02:10,578 adoption. You have to figure out how to make people actually see this 35 00:02:10,634 --> 00:02:14,324 change or want it and have it actually work for them. 36 00:02:14,492 --> 00:02:18,212 So how do we get there? It turns out we may be missing 37 00:02:18,276 --> 00:02:21,188 a few layers of the AI agent stack. 38 00:02:21,364 --> 00:02:24,436 Actually, we are missing three necessary layers right now, 39 00:02:24,508 --> 00:02:28,420 plus a bonus. The accountability layer the foundation of 40 00:02:28,460 --> 00:02:32,180 transparency. Verifiable work and reasoning 41 00:02:32,340 --> 00:02:35,876 the context layer, A system to unlock company 42 00:02:35,948 --> 00:02:38,530 knowledge, culture and goals. 43 00:02:39,030 --> 00:02:42,862 The coordination layer, enabling agents to collaborate 44 00:02:42,926 --> 00:02:47,214 seamlessly with shared knowledge systems Empowering 45 00:02:47,262 --> 00:02:50,702 AI agents, equipping them with the tools and software 46 00:02:50,766 --> 00:02:54,318 to maximize their autonomy in the rising B to a 47 00:02:54,374 --> 00:02:58,238 sphere. We are interested in companies building across each 48 00:02:58,294 --> 00:03:01,982 one of these layers or connecting them all, like NFX 49 00:03:02,046 --> 00:03:05,352 portfolio company Misa. More on that below. 50 00:03:05,536 --> 00:03:08,696 As we solve these challenges and build this infrastructure, 51 00:03:08,808 --> 00:03:12,888 we'll be able to tackle new and more complex and valuable tasks with 52 00:03:12,944 --> 00:03:16,712 AI. And once that's the norm, many more markets we 53 00:03:16,736 --> 00:03:20,440 can barely even conceive of now will emerge. But first 54 00:03:20,560 --> 00:03:24,664 we need these layers and. Here'S why Unlocking Autonomy 55 00:03:24,792 --> 00:03:28,088 from RPA to APA Agentic Process 56 00:03:28,144 --> 00:03:32,048 Automation to understand how. We are going to unlock full 57 00:03:32,104 --> 00:03:35,664 autonomy, we first have to understand a major shift 58 00:03:35,712 --> 00:03:38,672 in the way people look at process automation. 59 00:03:38,816 --> 00:03:42,560 For lack of a more interesting word, we are moving from 60 00:03:42,600 --> 00:03:46,272 robotic process automation to 61 00:03:46,296 --> 00:03:48,500 an agentic process automation. 62 00:03:51,240 --> 00:03:55,344 RPA is a multi billion dollar industry with massive 63 00:03:55,392 --> 00:03:59,478 companies like UiPath, Blue Prism and Workfusion among others. 64 00:03:59,664 --> 00:04:03,674 It's proof of concept that people are more than willing to adopt automation 65 00:04:03,722 --> 00:04:07,162 for high value tasks. To understand how we can bring 66 00:04:07,186 --> 00:04:10,890 on the agent economy, it's useful to use RPA as 67 00:04:10,930 --> 00:04:14,730 a starting point. Once you see its benefits and limitations, 68 00:04:14,890 --> 00:04:19,082 it's clear how agents are the natural and massive next step. 69 00:04:19,266 --> 00:04:23,338 The Benefits RPA excels at rule based structured 70 00:04:23,434 --> 00:04:27,076 tasks spanning multiple business systems 100 71 00:04:27,148 --> 00:04:30,996 to 200 steps. It was effective at capturing company 72 00:04:31,068 --> 00:04:35,252 knowledge within rules, for example VAT number processing, 73 00:04:35,396 --> 00:04:39,956 making automations reliable as long as underlying systems are static 74 00:04:40,148 --> 00:04:43,284 and RPA has strong product market fit 75 00:04:43,332 --> 00:04:47,012 already. The Limitations the universe of possible 76 00:04:47,116 --> 00:04:50,452 RPA able tasks was always going to be limited 77 00:04:50,596 --> 00:04:54,014 because you had to in detail be able map 78 00:04:54,062 --> 00:04:57,294 out exactly what process the RPA should take, 79 00:04:57,462 --> 00:05:00,558 move a mouse here, design this spreadsheet that way, 80 00:05:00,614 --> 00:05:04,222 etc. And importantly, expect it to remain the 81 00:05:04,246 --> 00:05:08,110 same or it breaks. RPA can only go so 82 00:05:08,150 --> 00:05:11,774 far because you can't process map and expect perfect 83 00:05:11,822 --> 00:05:15,566 exact repeatability in everything you do. Some companies 84 00:05:15,638 --> 00:05:19,582 can't even process map at all without hiring outside consultants 85 00:05:19,646 --> 00:05:23,582 to mine their own processes. In fact, 86 00:05:23,686 --> 00:05:27,102 you may not even want that dynamic all the time. Part of 87 00:05:27,126 --> 00:05:30,062 doing great work is reacting to an environment, 88 00:05:30,246 --> 00:05:33,850 intaking changes, tweaking things as you go. 89 00:05:34,230 --> 00:05:38,126 In summary, RPA works extremely well for certain tasks, 90 00:05:38,238 --> 00:05:41,022 but RPA is completely inflexible. 91 00:05:41,166 --> 00:05:45,170 Reliable but inflexible. Enter LLMs. 92 00:05:45,760 --> 00:05:48,824 The rise of LLMs represents a major shift. 93 00:05:48,952 --> 00:05:52,856 LLMs provide unlimited, cheap adaptive intelligence. 94 00:05:53,048 --> 00:05:56,424 They allowed us to define and collate the context needed 95 00:05:56,472 --> 00:06:00,024 to solve more complex problems. And as they began to learn 96 00:06:00,072 --> 00:06:03,480 reasoning, they hugely expanded the surface area of 97 00:06:03,520 --> 00:06:07,464 automatable tasks. That said, LLMs aren't perfect 98 00:06:07,512 --> 00:06:11,576 either. LLMs struggle with repetitive steps, but work well 99 00:06:11,648 --> 00:06:15,354 for unstructured parts of business processes. This can be a 100 00:06:15,362 --> 00:06:19,466 blessing or a curse depending on how creative versus deterministic 101 00:06:19,498 --> 00:06:22,954 you want your outcome to be. But either way, they're a 102 00:06:22,962 --> 00:06:26,730 black box. You can't be 100% sure of what the system is 103 00:06:26,770 --> 00:06:30,266 going to do, nor why it will do it. Even reasoning 104 00:06:30,298 --> 00:06:33,450 traces or model provided rationales can be 105 00:06:33,490 --> 00:06:36,618 completely hallucinated. Organizations need 106 00:06:36,674 --> 00:06:40,390 certainty or it's hard to implement any kind of system. 107 00:06:40,720 --> 00:06:43,784 Even if you want an LLM to be more creative, 108 00:06:43,912 --> 00:06:47,848 that's useless to you if you don't understand why and how 109 00:06:47,904 --> 00:06:51,688 it's arriving at certain conclusions. So where does this leave 110 00:06:51,704 --> 00:06:54,520 us? RPAs have strong PMF. 111 00:06:54,680 --> 00:06:57,432 It's easy to see how your system is working, 112 00:06:57,616 --> 00:07:02,136 but tasks are limited and they have no true flexibility or understanding 113 00:07:02,168 --> 00:07:06,180 of context. They also require a lot of pre work. 114 00:07:07,190 --> 00:07:11,326 LLMs are more capable with the unstructured information that's 115 00:07:11,358 --> 00:07:14,878 hard to express in rules, but they're a black box. 116 00:07:15,054 --> 00:07:18,142 The answer for agents and apa, we need a bit 117 00:07:18,166 --> 00:07:21,902 of both. We need the reliability of the RPA system with 118 00:07:21,926 --> 00:07:25,790 the flexibility and affordability of the LLM. 119 00:07:25,950 --> 00:07:29,582 This takes shape as an auditability and context layer 120 00:07:29,646 --> 00:07:33,374 that we can implement into the AI agent stack. As a 121 00:07:33,382 --> 00:07:36,782 builder in this space, you need to be working on this if you want to 122 00:07:36,806 --> 00:07:40,366 have a chance at widespread adoption. The Accountability 123 00:07:40,478 --> 00:07:44,530 Layer an unlock for Adoption, learning and Supervision 124 00:07:45,030 --> 00:07:48,110 Think back to your. Math classes in elementary school. 125 00:07:48,230 --> 00:07:51,806 When you were asked to solve a problem, you didn't get full credit 126 00:07:51,838 --> 00:07:56,046 for just writing the answer. You were asked to show your work. 127 00:07:56,198 --> 00:07:59,742 The teacher does this to verify that you actually understand 128 00:07:59,846 --> 00:08:03,960 the process that led to that correct answer. This is a 129 00:08:04,000 --> 00:08:07,592 step that many AI systems, even those that seem to show 130 00:08:07,616 --> 00:08:11,592 us logical trains of thought, are missing. We have no idea 131 00:08:11,656 --> 00:08:15,464 why AI actually generated those exact actions or chains 132 00:08:15,512 --> 00:08:19,480 of thought they're just generated. We first became aware of how 133 00:08:19,520 --> 00:08:23,656 big of a deal this was when we met Mesa. This metaphor 134 00:08:23,688 --> 00:08:27,412 was developed by David Villalon and Manuel Romero, 135 00:08:27,576 --> 00:08:30,764 the company's co founders, and it perfectly 136 00:08:30,812 --> 00:08:34,924 encapsulates the problem with so many AI agent ecosystems 137 00:08:34,972 --> 00:08:38,524 Right now, enterprises feel like they're supposed to blindly 138 00:08:38,572 --> 00:08:42,748 trust the AI's thought process. Early during product development, 139 00:08:42,884 --> 00:08:46,476 Misa met with a client that said they needed to prove exactly 140 00:08:46,508 --> 00:08:49,804 what was being done by their AI systems. For auditors, 141 00:08:49,932 --> 00:08:53,260 they needed evidence of each step taken and, critically, 142 00:08:53,340 --> 00:08:57,300 why those steps were taken at all. Conversations like that 143 00:08:57,340 --> 00:09:00,640 gave rise to Misa's concept of chain of work, 144 00:09:01,020 --> 00:09:04,404 a factor we now believe will be key to AI agent 145 00:09:04,452 --> 00:09:07,700 implementation in the workforce. At the heart 146 00:09:07,740 --> 00:09:11,320 of it sits MISA's knowledge processing unit, 147 00:09:12,780 --> 00:09:16,388 their proprietary reasoning engine for orchestrating each 148 00:09:16,444 --> 00:09:20,260 AI step as code rather than relying on ephemeral 149 00:09:20,340 --> 00:09:24,076 chain of thought text. By separating reasoning 150 00:09:24,108 --> 00:09:27,260 from execution they achieve deterministic, 151 00:09:27,340 --> 00:09:30,732 auditable outcomes. Every action is logged in 152 00:09:30,756 --> 00:09:33,900 an explicit chain of work, bridging the best 153 00:09:33,940 --> 00:09:37,308 of LLM style creativity with the reliability 154 00:09:37,404 --> 00:09:41,292 of traditional software. Unlike typical RPA or 155 00:09:41,316 --> 00:09:45,068 Frontier Lab solutions, which remain mostly guesswork behind 156 00:09:45,124 --> 00:09:48,240 the scenes, the KPU fosters trust. 157 00:09:48,620 --> 00:09:52,596 Teams can see precisely why and how the AI 158 00:09:52,708 --> 00:09:56,452 took each action, correct or refine any step and 159 00:09:56,476 --> 00:10:00,164 roll out changes consistently. I like to joke with founders 160 00:10:00,212 --> 00:10:03,812 that I work with that the best B2B software products are 161 00:10:03,836 --> 00:10:07,812 those that help people get promoted. Those that internal stakeholders 162 00:10:07,876 --> 00:10:11,396 smell that they can get big recognition by bringing it in. 163 00:10:11,548 --> 00:10:15,132 That's the reward that AI promises today, but it also comes 164 00:10:15,156 --> 00:10:18,796 with risk. No one wants to bring in a system that ultimately 165 00:10:18,828 --> 00:10:22,476 doesn't work. Building this accountability tips the risk 166 00:10:22,548 --> 00:10:25,884 reward ratio back into your favor. It's a 167 00:10:25,892 --> 00:10:29,840 given that AI automation is a huge win for enterprises. 168 00:10:30,260 --> 00:10:34,252 The key is reducing the risks, real and perceived, 169 00:10:34,396 --> 00:10:38,316 associated with implementation. Misa's chain 170 00:10:38,348 --> 00:10:41,730 of work helps with that. Ratio, and it's working 171 00:10:42,430 --> 00:10:46,518 the Context Layer what makes a great employee? 172 00:10:46,694 --> 00:10:49,990 What makes a great hire? It's not just the credentials. 173 00:10:50,070 --> 00:10:54,006 It's not just the experience. Ultimately, an employee's success 174 00:10:54,078 --> 00:10:57,158 in your organization will depend on their style, 175 00:10:57,254 --> 00:11:00,902 adaptability, and critically also on your ability 176 00:11:00,966 --> 00:11:04,450 to communicate what and how you want things to be done. 177 00:11:04,750 --> 00:11:08,230 Example, you hire a marketer who takes the time to 178 00:11:08,270 --> 00:11:12,280 understand your brand's voice and why you say what you need to say 179 00:11:12,400 --> 00:11:15,660 rather than just churning out bland marketing copy. 180 00:11:16,320 --> 00:11:19,944 Example, you hire an HR person that understands 181 00:11:19,992 --> 00:11:23,928 that he she is actually building company culture, not just 182 00:11:23,984 --> 00:11:28,632 creating an employee handbook. This is the key reason GPT4 183 00:11:28,736 --> 00:11:32,008 isn't an amazing employee. No matter what you do, 184 00:11:32,144 --> 00:11:35,762 GPT4 doesn't get you nor your company if it 185 00:11:35,786 --> 00:11:39,634 acts according to a set of rules, but it lacks the nuance and decision 186 00:11:39,722 --> 00:11:43,538 making context you'd expect from a human employee. Even if you 187 00:11:43,594 --> 00:11:47,234 were to articulate those rules to an AI workflow or custom 188 00:11:47,322 --> 00:11:51,282 GPT, you'd never get all of them. For a few reasons. 189 00:11:51,346 --> 00:11:55,234 A lot of what we learn at a new job isn't written down anywhere. 190 00:11:55,362 --> 00:11:59,122 It's learned by observation, intuition, through receiving 191 00:11:59,186 --> 00:12:02,992 feedback and asking clarifying questions. It's usually the 192 00:12:03,016 --> 00:12:06,512 ability to access and incorporate the unwritten stuff that 193 00:12:06,536 --> 00:12:09,460 distinguishes a great from a good employee. 194 00:12:10,040 --> 00:12:13,728 The actual stuff that is written is all in unstructured data, 195 00:12:13,864 --> 00:12:17,264 not in a database, but in PDFs with instructions, 196 00:12:17,392 --> 00:12:20,128 code, even in company emails. 197 00:12:20,304 --> 00:12:23,552 Most AI tools at the moment aren't plugged into the 198 00:12:23,576 --> 00:12:26,752 unstructured data ecosystem of a company let 199 00:12:26,776 --> 00:12:30,348 alone the minds of the current employees. We've talked about 200 00:12:30,404 --> 00:12:34,684 how one of the advantages of agents versus RPA is precisely 201 00:12:34,732 --> 00:12:38,492 this contextual understanding. It provides adaptability 202 00:12:38,636 --> 00:12:42,800 and eliminates the need for insanely costly process mapping. 203 00:12:43,220 --> 00:12:46,844 Organizing this knowledge is possible and it's been proven 204 00:12:46,892 --> 00:12:50,684 in more constrained environments. Industry standard 205 00:12:50,732 --> 00:12:55,132 retrieval, augmented generation are 206 00:12:55,156 --> 00:12:58,876 a decent start, but they eventually break under large data 207 00:12:58,948 --> 00:13:02,652 sets or specialized knowledge, making this a challenge. 208 00:13:02,796 --> 00:13:06,700 Misa approaches this differently by developing a virtual context 209 00:13:06,780 --> 00:13:10,476 window. VCW bypasses these complexities 210 00:13:10,588 --> 00:13:13,800 by functioning as an OS like paging system. 211 00:13:14,180 --> 00:13:17,900 Digital workers load and navigate only the data they 212 00:13:17,940 --> 00:13:22,076 need per step, giving them effectively unlimited memory and zero 213 00:13:22,148 --> 00:13:25,932 collisions. No fine tuning or unwieldy indexes 214 00:13:25,996 --> 00:13:29,516 needed. Crucially, the VCW also doubles 215 00:13:29,548 --> 00:13:33,516 as a long term know how store for each worker, meaning they adapt 216 00:13:33,548 --> 00:13:36,720 to new instructions or data seamlessly. 217 00:13:37,220 --> 00:13:41,004 A critical part of the AI agent stack must be this contextual 218 00:13:41,052 --> 00:13:44,972 layer. Your customer will think of this as space where they onboard an 219 00:13:44,996 --> 00:13:48,748 AI worker into their organization's unique approach and style. 220 00:13:48,924 --> 00:13:53,084 The challenge is to devise a way to encapsulate that context 221 00:13:53,132 --> 00:13:56,876 for your customers and translate that into your agent's 222 00:13:56,908 --> 00:14:00,892 DNA, both at the moment of onboarding and in the future, 223 00:14:01,076 --> 00:14:04,828 enabling usage of that knowledge and continuous learning. 224 00:14:05,004 --> 00:14:08,892 Some other initiatives in this broader area we have seen unstructured 225 00:14:08,956 --> 00:14:11,612 data preparation for AI agents, 226 00:14:11,756 --> 00:14:15,052 continuous systems to gather and generate new context 227 00:14:15,116 --> 00:14:18,800 data systems that allow us to fine tune models more 228 00:14:18,840 --> 00:14:22,704 easily. Memory systems and long context windows. 229 00:14:22,832 --> 00:14:26,912 See one of the latest advancements here, AI with 230 00:14:26,936 --> 00:14:30,656 an intuitive understanding of emotional intelligence and personality 231 00:14:30,768 --> 00:14:34,208 which will help with all of the above. See our 232 00:14:34,264 --> 00:14:38,288 piece Software with a Soul, the Coordination 233 00:14:38,384 --> 00:14:41,776 layer. Managing the Agentic Workforce in the future. 234 00:14:41,848 --> 00:14:45,312 Businesses are probably going to manage a set of AI agent 235 00:14:45,376 --> 00:14:48,822 employees. You'll have agents for customer service, 236 00:14:49,006 --> 00:14:52,486 sales, HR accounting, and it's 237 00:14:52,518 --> 00:14:56,230 likely that different companies will provide each of these workforces. 238 00:14:56,390 --> 00:15:00,582 It's already starting to happen. We're seeing job listings for AI 239 00:15:00,646 --> 00:15:03,958 agents in the wild. Those agents will have to talk 240 00:15:04,014 --> 00:15:07,254 to humans and to each other. Those agents 241 00:15:07,302 --> 00:15:11,878 will also require permissioning and rules with important considerations 242 00:15:11,974 --> 00:15:15,750 for privacy and security. This is an interesting crux 243 00:15:15,830 --> 00:15:18,930 moment in the development of the AI agent space. 244 00:15:19,230 --> 00:15:22,902 It seems obvious that we will have swarms of agents speaking to 245 00:15:22,926 --> 00:15:26,022 one another, but you could imagine a world where that 246 00:15:26,046 --> 00:15:29,446 isn't the case. You could see a dynamic where companies, 247 00:15:29,598 --> 00:15:32,870 likely incumbents, look to own the whole system of 248 00:15:32,910 --> 00:15:36,822 agent building and managing. In that case, they would probably look 249 00:15:36,846 --> 00:15:39,716 to discourage collaboration with other systems. 250 00:15:39,838 --> 00:15:43,304 A winner take all dynamic. That said, there's Not a 251 00:15:43,312 --> 00:15:46,440 ton of evidence to suggest any AI products have 252 00:15:46,480 --> 00:15:50,584 developed that way so far. With the exception of GPUs, 253 00:15:50,712 --> 00:15:54,968 most of the raw materials needed to build AI products and systems like 254 00:15:55,024 --> 00:15:58,580 foundational models aren't owned by one or two companies. 255 00:15:58,880 --> 00:16:02,056 We have OpenAI, Claude, Gemini, 256 00:16:02,168 --> 00:16:05,102 Mistral, and now Deepseek. 257 00:16:05,296 --> 00:16:08,626 With the sheer number of startups we're seeing in the agent space 258 00:16:08,698 --> 00:16:12,914 right now, it seems more likely that someone deep in the AI agent 259 00:16:12,962 --> 00:16:16,706 world will solve the communications and permissioning problem 260 00:16:16,858 --> 00:16:20,546 faster than an incumbent can shut them out. Ultimately, 261 00:16:20,658 --> 00:16:24,270 a thriving agent ecosystem is a win win for everyone. 262 00:16:24,810 --> 00:16:28,594 From the customer perspective, it provides you with an endless pool 263 00:16:28,642 --> 00:16:32,706 of potential AI talent and the ability to choose the best fit 264 00:16:32,738 --> 00:16:36,518 for you. From a founder's perspective, it opens 265 00:16:36,534 --> 00:16:39,814 the door to network effects. Each new agent 266 00:16:39,862 --> 00:16:43,542 that's added to the ecosystem actually benefits you if 267 00:16:43,566 --> 00:16:47,078 you are the one facilitating the connections. In that 268 00:16:47,134 --> 00:16:50,406 case, interagent communication is essential. 269 00:16:50,598 --> 00:16:54,006 Companies on the forefront of this wave already understand this 270 00:16:54,078 --> 00:16:56,646 and are building multimodal capabilities. 271 00:16:56,838 --> 00:17:01,616 Mesa's KPU from, for example, is model agnostic. 272 00:17:01,808 --> 00:17:05,712 In a world where foundational models are continuously improving, 273 00:17:05,856 --> 00:17:09,088 flexibility is essential. But we will also need 274 00:17:09,144 --> 00:17:12,800 systems for agents to safely exchange and share knowledge. 275 00:17:12,960 --> 00:17:16,624 This is something to be thinking about now as all these agent 276 00:17:16,672 --> 00:17:20,432 ecosystems get up and running. The frontier 277 00:17:20,576 --> 00:17:24,384 Giving AI agents Tools for. The Job Once we tackle 278 00:17:24,432 --> 00:17:27,842 accountability, context and coordination, we get 279 00:17:27,866 --> 00:17:31,474 to the fun stuff. We're already seeing a market emerge 280 00:17:31,522 --> 00:17:35,138 for tools for AI agents, software that will make them better 281 00:17:35,194 --> 00:17:39,826 at their jobs. Some are calling this nascent space B2A 282 00:17:39,978 --> 00:17:43,282 business to agent. This will be a major unlock that 283 00:17:43,306 --> 00:17:46,674 takes agents from rank and file workers to autonomous 284 00:17:46,722 --> 00:17:50,178 decision makers. Imagine if humans weren't allowed to use 285 00:17:50,234 --> 00:17:53,630 calculators or computers. Once you deploy an agent, 286 00:17:53,710 --> 00:17:57,614 you have to set them up for success. We're already at the beginning of 287 00:17:57,622 --> 00:18:01,822 this world. We've seen chat GPT use a web browser, 288 00:18:01,966 --> 00:18:05,022 Claude move a cursor around a screen. 289 00:18:05,166 --> 00:18:08,830 ElevenLabs can give them a voice, but we can imagine this 290 00:18:08,870 --> 00:18:12,590 world getting 10 times better. Agents will need to be able 291 00:18:12,630 --> 00:18:16,094 to pay one another for services. They'll need to be able to 292 00:18:16,102 --> 00:18:20,374 enter into contracts or plug into systems where humans and programs 293 00:18:20,422 --> 00:18:24,182 already interact. Apps can inspire infrastructure 294 00:18:24,246 --> 00:18:28,390 and vice versa. Within the AI agent space, 295 00:18:28,510 --> 00:18:32,566 we're seeing this dynamic as well. These infrastructural layers 296 00:18:32,598 --> 00:18:35,878 will inspire apps, new types of agents, 297 00:18:35,974 --> 00:18:39,414 plus tools for agents which will inform progress at 298 00:18:39,422 --> 00:18:43,414 the infrastructure layer. Creating tools where agents themselves are 299 00:18:43,422 --> 00:18:46,914 the end user is a massive area of white space. 300 00:18:47,082 --> 00:18:51,410 We're watching it closely what it takes to onboard AI 301 00:18:51,490 --> 00:18:54,946 agents. Let's be clear. We are all in 302 00:18:55,018 --> 00:18:59,362 on agents and excited about the potential they hold to us 303 00:18:59,466 --> 00:19:02,882 and most of the founders we work with. The world where we are 304 00:19:02,906 --> 00:19:06,870 all using AI agents each day is an inevitability. 305 00:19:07,370 --> 00:19:11,122 Part of this excitement is building this new ecosystem 306 00:19:11,186 --> 00:19:14,814 from the bottom up. We have to really understand what it takes 307 00:19:14,862 --> 00:19:18,478 to get people to adopt a whole new computing paradigm. 308 00:19:18,654 --> 00:19:21,982 There is a life cycle to these things and we're only at 309 00:19:22,006 --> 00:19:25,582 the beginning. Creating these layers will be key to 310 00:19:25,606 --> 00:19:29,166 making AI agents tools that most people trust and 311 00:19:29,238 --> 00:19:33,070 use every day. These are the challenges that will catapult us over 312 00:19:33,110 --> 00:19:37,102 the adoption gap. We're excited for the companies that recognize 313 00:19:37,166 --> 00:19:40,982 this challenge and dove right in there the new infrastructure 314 00:19:41,046 --> 00:19:43,990 upon which the AI agent revolution will be built.