WEBVTT 00:00:00.000 --> 00:00:01.960 All right. Let's let's dive into something pretty 00:00:01.960 --> 00:00:05.019 fascinating today. We're talking about AI agents, 00:00:05.240 --> 00:00:07.059 you know, those AI systems that can actually 00:00:07.059 --> 00:00:10.660 do things out there. They can plan, use tools, 00:00:10.980 --> 00:00:13.500 maybe even team up. Right. And we've been looking 00:00:13.500 --> 00:00:15.140 through these sources articles, some research 00:00:15.140 --> 00:00:19.699 notes focusing on protocols, really, and some 00:00:19.699 --> 00:00:22.019 breakthroughs with these agents. Yeah. And what 00:00:22.019 --> 00:00:24.519 kind of jumps out right away is this this problem 00:00:24.519 --> 00:00:27.140 fragmentation. You've got these powerful agents, 00:00:27.339 --> 00:00:31.059 but. It feels like computers before the Internet 00:00:31.059 --> 00:00:34.679 had rules, you know, before TCP IP. Exactly. 00:00:34.799 --> 00:00:36.100 They can't really talk to each other easily. 00:00:36.219 --> 00:00:38.939 It's a bottleneck. A big one. So that's our mission 00:00:38.939 --> 00:00:41.840 for this deep dive. Figure out what this fragmentation 00:00:41.840 --> 00:00:44.039 actually means, how people are trying to fix 00:00:44.039 --> 00:00:47.119 it with these new protocols, and why this is 00:00:47.119 --> 00:00:49.140 just so crucial for where AI is going. We're 00:00:49.140 --> 00:00:50.960 even going to look at a really cool specific 00:00:50.960 --> 00:00:54.039 example later. Sounds good. It's definitely a 00:00:54.039 --> 00:00:56.219 key area right now. So diving in, the sources 00:00:56.219 --> 00:00:59.100 are clear. Right now, most AI agents are kind 00:00:59.100 --> 00:01:02.460 of isolated, siloed. That's the word they use, 00:01:02.520 --> 00:01:04.480 yeah, siloed. They have their own ways to connect 00:01:04.480 --> 00:01:08.719 to tools, specific APIs, custom wrappers. It's 00:01:08.719 --> 00:01:12.329 all a bit bespoke. Yeah, think of it like everyone's 00:01:12.329 --> 00:01:13.590 speaking a totally different language. You've 00:01:13.590 --> 00:01:16.069 got these really smart systems, right? But they're 00:01:16.069 --> 00:01:18.430 locked in their own worlds. You can't easily 00:01:18.430 --> 00:01:21.450 scale up complex tasks that need different types 00:01:21.450 --> 00:01:23.430 of agents working together. Just because they 00:01:23.430 --> 00:01:26.370 don't have a shared way to communicate. Or understand 00:01:26.370 --> 00:01:28.989 each other. Exactly. It's just inefficient. This 00:01:28.989 --> 00:01:32.129 fractured ecosystem, it's a real drag on progress, 00:01:32.390 --> 00:01:34.709 a big bottleneck. Okay. So if they're all speaking 00:01:34.709 --> 00:01:37.450 different languages, how do you fix that? What 00:01:37.450 --> 00:01:39.290 are the sources suggesting? Well, the strong 00:01:39.290 --> 00:01:42.189 consensus looking through this material is standard 00:01:42.189 --> 00:01:45.069 protocols. Standard protocols. Okay. It's really 00:01:45.069 --> 00:01:47.870 seen as the missing infrastructure layer. Just 00:01:47.870 --> 00:01:51.290 like the internet needed TCP, IP, HTTP, all that. 00:01:51.329 --> 00:01:53.870 Right. To let everything connect. Yeah. Agents 00:01:53.870 --> 00:01:56.420 need that too. Common language, common rules, 00:01:56.659 --> 00:01:59.519 how to find each other, talk securely, know what 00:01:59.519 --> 00:02:01.480 the other agent can actually do across different 00:02:01.480 --> 00:02:04.219 companies, different platforms. So it's like 00:02:04.219 --> 00:02:07.379 building the roads and maybe the traffic rules 00:02:07.379 --> 00:02:09.699 for this AI world. That's a great way to put 00:02:09.699 --> 00:02:12.800 it. Exactly. These protocols, they're the missing 00:02:12.800 --> 00:02:16.379 link between having powerful individual agents 00:02:16.379 --> 00:02:20.199 and enabling, you know, true large scale collaboration, 00:02:20.500 --> 00:02:23.169 working together effectively. And without that, 00:02:23.310 --> 00:02:26.289 multi -agent systems just don't really take off. 00:02:26.430 --> 00:02:28.729 Pretty much. Their potential is really capped. 00:02:28.830 --> 00:02:31.449 The sources mention one specific protocol that 00:02:31.449 --> 00:02:34.169 seems quite developed already, Anthropix Model 00:02:34.169 --> 00:02:38.289 Context Protocol. MCP. MCP. Okay, what's that 00:02:38.289 --> 00:02:41.409 one focused on? MCP is mostly about agents talking 00:02:41.409 --> 00:02:44.030 securely to tools or external resources. Ah, 00:02:44.150 --> 00:02:46.669 okay. Tools, not other agents. Right. It handles 00:02:46.669 --> 00:02:49.030 things like, is the agent using the tool correctly? 00:02:49.250 --> 00:02:51.270 What about data privacy when it hits an external 00:02:51.270 --> 00:02:53.969 system, managing the technical side of calling 00:02:53.969 --> 00:02:56.469 different APIs properly? So it's like the rulebook 00:02:56.469 --> 00:02:59.469 for an agent using a hammer or database or something? 00:02:59.650 --> 00:03:02.250 Kind of, yeah. A secure handshake, making sure 00:03:02.250 --> 00:03:04.719 that interaction works reliably and safely. Okay, 00:03:04.719 --> 00:03:07.020 that makes sense. Agents need to use tools. But 00:03:07.020 --> 00:03:09.300 what about agents talking to other agents? That 00:03:09.300 --> 00:03:11.840 feels even more like the core collaboration piece, 00:03:12.039 --> 00:03:13.580 right? If one agent needs help from another. 00:03:13.759 --> 00:03:15.500 Absolutely. You nailed it. That's the whole next 00:03:15.500 --> 00:03:19.020 level. Agent -to -agent communication. A2A. And 00:03:19.020 --> 00:03:21.300 the sources discuss several protocols being worked 00:03:21.300 --> 00:03:24.180 on specifically for that A2A interaction. Okay, 00:03:24.219 --> 00:03:27.020 so MCP for tools, A2A for agent -to -agent chats. 00:03:27.319 --> 00:03:30.000 Got it. Right. And even within A2A, there are 00:03:30.000 --> 00:03:32.280 like different flavors. The material mentions 00:03:32.280 --> 00:03:34.340 one protocol actually called A2A, which came 00:03:34.340 --> 00:03:36.960 out of Google. This one seems more geared towards 00:03:36.960 --> 00:03:40.379 internal use. Yeah. Like within one company. 00:03:40.560 --> 00:03:42.060 How does that work? It uses something called 00:03:42.060 --> 00:03:45.120 agent cards. Agent cards, like baseball cards. 00:03:45.360 --> 00:03:47.919 Kind of, yeah. More like a digital profile, usually 00:03:47.919 --> 00:03:51.159 JSON. An agent publishes its card saying, here's 00:03:51.159 --> 00:03:53.400 what I do. Here's the data I need. Talk to me 00:03:53.400 --> 00:03:55.969 like this. Okay. Then other agents within the 00:03:55.969 --> 00:03:57.930 company can look it up, find the right agent 00:03:57.930 --> 00:04:00.849 for a task, see its card, and know how to delegate 00:04:00.849 --> 00:04:03.710 stuff to it. Good for managing your own internal 00:04:03.710 --> 00:04:06.650 zoo of agents. Right. Makes sense for a big organization. 00:04:06.689 --> 00:04:09.530 Keep things tidy internally. Exactly. But then 00:04:09.530 --> 00:04:11.889 there's another one mentioned, the Agent Network 00:04:11.889 --> 00:04:15.669 Protocol, AMP. AMP, different. Yeah. This one's 00:04:15.669 --> 00:04:18.269 open source, and it's aiming bigger. It's designed 00:04:18.269 --> 00:04:20.189 for agents talking across different organizations. 00:04:20.860 --> 00:04:23.319 Different companies, labs, whatever. Ah, so more 00:04:23.319 --> 00:04:25.699 like the public internet, less like a company 00:04:25.699 --> 00:04:28.019 intranet. That's a really good analogy, actually. 00:04:28.339 --> 00:04:32.000 A2A, Google's one, is like the corporate intranet 00:04:32.000 --> 00:04:34.759 for agents. ANP is trying to be the public internet 00:04:34.759 --> 00:04:36.839 for agents. Yeah. Decentralized collaboration. 00:04:37.439 --> 00:04:39.439 Agents that maybe don't know each other beforehand. 00:04:39.759 --> 00:04:41.879 Okay, that's a key difference. Internal versus 00:04:41.879 --> 00:04:44.459 external. Precisely. And the sources toss out 00:04:44.459 --> 00:04:46.259 a few other names, too, just showing how active 00:04:46.259 --> 00:04:49.199 this space is. There's Agora. which sounds interesting. 00:04:49.259 --> 00:04:51.720 It's more user -centric. Agents kind of figure 00:04:51.720 --> 00:04:53.980 out how to talk using these protocol documents. 00:04:54.439 --> 00:04:58.199 Maybe even in natural language, less rigid. And 00:04:58.199 --> 00:05:00.720 then you've got really specific ones, domain 00:05:00.720 --> 00:05:04.199 -specific, like CrowdES or SPPs. Those are for 00:05:04.199 --> 00:05:06.160 agents controlling robots, like in a warehouse 00:05:06.160 --> 00:05:09.899 or something. Or PXP and LKA, which are focused 00:05:09.899 --> 00:05:12.360 on agents helping humans directly. So they need 00:05:12.360 --> 00:05:15.240 good human -agent interaction rules. Okay, so 00:05:15.240 --> 00:05:18.779 it's not just one answer. It's a whole... ecosystem 00:05:18.779 --> 00:05:22.019 of communication needs. Exactly. Different protocols 00:05:22.019 --> 00:05:24.319 for different kinds of agent teamwork. Man, that's 00:05:24.319 --> 00:05:28.699 a lot of acronyms. MCP, A2A, ANP, Agora. How 00:05:28.699 --> 00:05:31.319 do people even compare these? How do you know 00:05:31.319 --> 00:05:34.319 which one is good or right for a job? Yeah, that's 00:05:34.319 --> 00:05:36.860 a really important question. And the sources 00:05:36.860 --> 00:05:39.240 actually get into this. One paper lays out specific 00:05:39.240 --> 00:05:41.519 criteria for evaluating these protocols. Okay, 00:05:41.540 --> 00:05:43.660 like what? Well, the big ones they mention are 00:05:43.660 --> 00:05:46.259 security, obviously. Is the communication safe? 00:05:46.579 --> 00:05:49.439 Private. especially with sensitive data flying 00:05:49.439 --> 00:05:52.480 around crucial yeah then operability basically 00:05:52.480 --> 00:05:54.779 how easy is it for developers to actually use 00:05:54.779 --> 00:05:57.500 this protocol is it clear well documented you 00:05:57.500 --> 00:05:59.040 don't want it to be a huge pain to implement 00:05:59.040 --> 00:06:02.620 needs to be practical extensibility ai changes 00:06:02.620 --> 00:06:06.000 fast yeah so can the protocol adapt can you add 00:06:06.000 --> 00:06:08.420 new features support new kinds of agents or tools 00:06:08.420 --> 00:06:11.360 later on without breaking everything future proofing 00:06:11.360 --> 00:06:14.800 kind of exactly trustworthiness Is it reliable? 00:06:14.959 --> 00:06:17.639 Does it work consistently? Can you trust that 00:06:17.639 --> 00:06:20.100 agent interactions will happen as expected? Makes 00:06:20.100 --> 00:06:23.120 sense. And finally, integration capacity. How 00:06:23.120 --> 00:06:25.459 well does it connect different things? Agents 00:06:25.459 --> 00:06:27.560 on different platforms, maybe cloud versus local 00:06:27.560 --> 00:06:30.560 device. Can it bridge those gaps? Okay, so it's 00:06:30.560 --> 00:06:33.079 way more than just can they send messages. It's 00:06:33.079 --> 00:06:35.519 about making the whole interaction solid, secure, 00:06:35.839 --> 00:06:38.399 usable, adaptable. Exactly. It's about building 00:06:38.399 --> 00:06:41.779 a robust communication system. These criteria... 00:06:42.009 --> 00:06:45.550 are how the community is sort of weighing these 00:06:45.550 --> 00:06:46.970 different approaches, figuring out what works 00:06:46.970 --> 00:06:49.310 best, where it's complex, but super important 00:06:49.310 --> 00:06:51.350 work. Okay, this is definitely getting technical. 00:06:51.470 --> 00:06:52.709 And you might be listening and thinking, all 00:06:52.709 --> 00:06:55.850 right, protocols, standards, why should I care? 00:06:55.990 --> 00:06:57.949 How does this affect me? It's a fair point. It 00:06:57.949 --> 00:07:00.529 sounds like plumbing, maybe. But the sources 00:07:00.529 --> 00:07:03.110 really emphasize that these AI agents are getting 00:07:03.110 --> 00:07:04.930 really capable. They're moving out of the lab. 00:07:05.009 --> 00:07:07.769 They could become like basic infrastructure in 00:07:07.769 --> 00:07:10.790 business, science, maybe even our homes eventually. 00:07:11.290 --> 00:07:13.029 Everywhere. Yeah, it doesn't feel like sci -fi 00:07:13.029 --> 00:07:15.509 anymore. It feels closer. Right. And if they 00:07:15.509 --> 00:07:17.610 stay fragmented, stuck in those silos we talked 00:07:17.610 --> 00:07:21.709 about, their potential is just limited. Massively 00:07:21.709 --> 00:07:23.970 limited. Imagine trying to run, I don't know, 00:07:24.009 --> 00:07:26.490 global shipping if every port used a completely 00:07:26.490 --> 00:07:29.129 different incompatible system for manifests. 00:07:29.529 --> 00:07:32.319 It'd be chaos. Yeah. Utterly inefficient. Yeah, 00:07:32.360 --> 00:07:34.620 nothing would work together. So solving this 00:07:34.620 --> 00:07:37.379 fragmentation with protocols, it's not just about 00:07:37.379 --> 00:07:40.480 neat code. It's fundamental to unlocking what 00:07:40.480 --> 00:07:42.620 these agents can do. It lets them team up on 00:07:42.620 --> 00:07:45.819 big problems, share info securely, automate whole 00:07:45.819 --> 00:07:48.819 workflows across different systems. That's where 00:07:48.819 --> 00:07:50.639 the real value is going to come from. So they 00:07:50.639 --> 00:07:52.660 become more than just clever tools. They become 00:07:52.660 --> 00:07:55.420 part of a bigger coordinated system. Exactly. 00:07:55.980 --> 00:07:58.360 Whether that's making businesses run smoother, 00:07:58.500 --> 00:08:00.439 speeding up research by connecting different 00:08:00.439 --> 00:08:03.220 AI analysis tools, or even just making our personal 00:08:03.220 --> 00:08:05.839 tech work together better for us. Without protocols, 00:08:06.100 --> 00:08:08.439 they're like brilliant specialists who just can't 00:08:08.439 --> 00:08:11.420 coordinate on a complex surgery. With protocols, 00:08:11.759 --> 00:08:14.180 they can potentially form that expert surgical 00:08:14.180 --> 00:08:17.160 team. Okay, that makes the stakes clearer. It's 00:08:17.160 --> 00:08:19.279 about enabling that next level of capability. 00:08:19.680 --> 00:08:22.060 Definitely. And speaking of capability, let's 00:08:22.060 --> 00:08:24.000 switch gears to that concrete example you mentioned. 00:08:24.569 --> 00:08:26.569 The sources highlight this really interesting 00:08:26.569 --> 00:08:29.750 breakthrough with an autonomous AI agent in medicine. 00:08:29.970 --> 00:08:31.689 Yeah, this sounded pretty wild. What did it do? 00:08:31.910 --> 00:08:34.129 So this was research from TU Dresden and collaborators. 00:08:34.570 --> 00:08:36.450 They built an agent designed to help doctors 00:08:36.450 --> 00:08:39.230 with clinical decisions in oncology. Cancer care. 00:08:39.429 --> 00:08:42.409 Okay. High stakes. Very. And in their early tests, 00:08:42.610 --> 00:08:44.830 using simulated but realistic patient cases, 00:08:45.049 --> 00:08:48.450 this agent hit 91 % accuracy. Whoa, hold on. 00:08:48.529 --> 00:08:51.889 91 % accuracy on recommending cancer treatments 00:08:51.889 --> 00:08:54.620 or diagnoses on simulated cases. Yeah. based 00:08:54.620 --> 00:08:56.879 on the simulated patient data provided. Wow, 00:08:57.460 --> 00:09:00.460 that seems incredibly high. It is. And crucially, 00:09:00.559 --> 00:09:02.480 the sources point out, it wasn't just accurate, 00:09:02.639 --> 00:09:05.659 it was also grounded. It correctly cited official 00:09:05.659 --> 00:09:09.139 treatment guidelines 75 % of the time. Ah, so 00:09:09.139 --> 00:09:11.679 it showed its work, basically, based on actual 00:09:11.679 --> 00:09:14.340 medical standards. Exactly, which is vital in 00:09:14.340 --> 00:09:16.080 medicine. You can't just have a black box making 00:09:16.080 --> 00:09:18.480 recommendations. So how did they build this thing? 00:09:18.539 --> 00:09:22.019 Is it just a standard LLM? It's built on GPT 00:09:22.019 --> 00:09:23.860 -4, but it's way more than that. They give it 00:09:23.860 --> 00:09:26.940 specialized tools. It can analyze medical images, 00:09:27.259 --> 00:09:29.820 you know, MRI scans, CT scans. It can help graph 00:09:29.820 --> 00:09:32.440 radiology reports. It can even predict genetic 00:09:32.440 --> 00:09:34.919 mutations from molecular data. OK, so it has 00:09:34.919 --> 00:09:37.259 specialist skills. Right. And it uses search 00:09:37.259 --> 00:09:40.279 tools, PubMed for research papers, OncoKB for 00:09:40.279 --> 00:09:42.840 cancer knowledge, Google for general info to 00:09:42.840 --> 00:09:44.840 pull in the latest evidence. It keeps itself 00:09:44.840 --> 00:09:48.120 up to date. So GPT -4 plus a whole medical toolkit 00:09:48.120 --> 00:09:50.889 and library access. That's a good summary. And 00:09:50.889 --> 00:09:52.750 critically, the training wasn't just random web 00:09:52.750 --> 00:09:55.830 data. They fed it over 6 ,800 official oncology 00:09:55.830 --> 00:09:59.389 documents, protocols, trial data, guidelines. 00:09:59.769 --> 00:10:01.570 So it learned from the best available medical 00:10:01.570 --> 00:10:04.570 knowledge? Precisely. That grounding in high 00:10:04.570 --> 00:10:07.429 -quality data is likely key to its performance. 00:10:07.769 --> 00:10:10.750 And the testing, that 91%, was it on easy stuff? 00:10:11.320 --> 00:10:13.940 Nope. They stress it was on 20 complex, realistic 00:10:13.940 --> 00:10:16.440 cancer scenarios designed to be challenging. 00:10:16.659 --> 00:10:19.240 And human experts checked the agent's outputs 00:10:19.240 --> 00:10:21.940 for accuracy and made sure the citations were 00:10:21.940 --> 00:10:25.000 relevant. Hmm. OK. 20 cases isn't thousands. 00:10:25.059 --> 00:10:28.379 But still, that's genuinely impressive potential. 00:10:28.419 --> 00:10:30.580 What are the implications? Well, the researchers 00:10:30.580 --> 00:10:32.899 are cautious. It's early stage, obviously. Needs 00:10:32.899 --> 00:10:35.120 way more testing, validation. But the potential 00:10:35.120 --> 00:10:37.909 implications are huge. If this holds up. And 00:10:37.909 --> 00:10:39.850 with the right safeguards and absolutely critical 00:10:39.850 --> 00:10:42.769 human oversight, this is decision support, not 00:10:42.769 --> 00:10:44.750 replacement. Right. Assisting the doctor, not 00:10:44.750 --> 00:10:47.090 being the doctor. Exactly. This kind of aging 00:10:47.090 --> 00:10:49.409 could become a standard tool. Helping oncologists 00:10:49.409 --> 00:10:51.789 manage the information overload, consider all 00:10:51.789 --> 00:10:53.889 the latest evidence, maybe improve consistency 00:10:53.889 --> 00:10:56.289 in care. And you can imagine this extending to 00:10:56.289 --> 00:10:58.620 other complex medical fields, too. Okay, see 00:10:58.620 --> 00:11:01.059 that example really brings the power home. An 00:11:01.059 --> 00:11:04.659 agent capable of that kind of analysis. Now imagine 00:11:04.659 --> 00:11:07.200 that agent being able to seamlessly talk to another 00:11:07.200 --> 00:11:10.360 agent that manages patient scheduling, or one 00:11:10.360 --> 00:11:14.559 that finds clinical trials, or one handling insurance 00:11:14.559 --> 00:11:18.700 paperwork, all using a shared protocol like ANP 00:11:18.700 --> 00:11:22.100 or something similar, reliably, securely. That's 00:11:22.100 --> 00:11:23.720 the future of these protocols unlock, right? 00:11:23.740 --> 00:11:25.759 That's exactly it. You have these agents getting 00:11:25.759 --> 00:11:29.039 incredibly good at specific complex tasks like 00:11:29.039 --> 00:11:32.100 that medical analysis. But for AI to truly scale, 00:11:32.159 --> 00:11:34.419 to weave itself into how things actually get 00:11:34.419 --> 00:11:37.000 done, we absolutely have to solve that fundamental 00:11:37.000 --> 00:11:38.840 communication problem. How do they find each 00:11:38.840 --> 00:11:40.620 other? How do they talk securely? How do they 00:11:40.620 --> 00:11:43.159 collaborate reliably? Protocols are that missing 00:11:43.159 --> 00:11:45.179 foundation needed to build that interconnected 00:11:45.179 --> 00:11:48.460 AI future. Right. So wrapping this up then, we've 00:11:48.460 --> 00:11:50.639 seen the problem. AI agents kind of stuck in 00:11:50.639 --> 00:11:52.980 their own worlds, fragmented. We've seen the 00:11:52.980 --> 00:11:55.480 potential solutions taking shape. These different 00:11:55.480 --> 00:11:58.440 protocols, MCP for tools, A2A and AMP for agent 00:11:58.440 --> 00:12:00.940 -to -agent comms, each with its own focus. Yeah, 00:12:00.980 --> 00:12:03.500 the infrastructure being built. And we got a 00:12:03.500 --> 00:12:05.840 glimpse of the incredible potential with that 00:12:05.840 --> 00:12:07.860 medical agent example showing what they can do 00:12:07.860 --> 00:12:09.779 individually and hinting at what they could do 00:12:09.779 --> 00:12:11.600 together. Right. And maybe the final thought 00:12:11.600 --> 00:12:14.080 for you listening is this. Considering how fast 00:12:14.080 --> 00:12:16.419 these agents are evolving, and how much we need 00:12:16.419 --> 00:12:19.860 them to coordinate for complex tasks. Ask yourself, 00:12:20.039 --> 00:12:23.659 as AI agents get woven more into our world, how 00:12:23.659 --> 00:12:25.879 will their ability or maybe their inability to 00:12:25.879 --> 00:12:28.539 communicate effectively reshape, well, everything, 00:12:28.679 --> 00:12:31.100 the industries, the systems, maybe even our daily 00:12:31.100 --> 00:12:33.539 lives? It's definitely something worth thinking 00:12:33.539 --> 00:12:33.799 about.