0:02 Traditional software follows a rhythm we 0:04 all understand. Input, process, output. 0:07 Predictable, reliable, like clockwork. 0:10 But imagine if these years couldn't 0:12 think for themselves. If they could 0:14 decide when to turn, how fast to spin, 0:17 and in which direction to go. Hi. Uh, 0:20 thank you for having me. You want me to 0:22 go? Traditional code is like a script 0:24 for actors. Each line tells the program 0:26 exactly what to do and when to do it and 0:28 how to do it. Believe me when I tell you 0:31 that this is over. That that was good. 0:34 Oh yeah. Okay. I heard that there were 0:36 some an AI agent is like an improvising 0:39 actor. You give it a character 0:41 description and an objective, but it 0:43 decides how to get there. The difference 0:45 isn't just flexibility. It is 0:47 understanding purpose over procedure. 0:52 I thought I 0:53 understood. Just I'm so 0:57 sorry. Please, please let this not be 1:00 over. I thank 1:02 you. Okay. Does that Does that work? 1:05 Yeah, that was that was amazing. Thank 1:07 you. Okay. At the heart of every agent, 1:10 there's a decision loop that might feel 1:11 familiar. This isn't new to you. You use 1:14 this cycle every time you navigate 1:16 traffic, cook dinner, or solve a problem 1:18 at work. What has changed is that we 1:20 have encoded this loop into systems that 1:22 can continuously evaluate and respond. 1:25 When we program, we usually dictate 1:27 exactly when each tool gets used. An 1:29 agent selects its own tools based on the 1:32 situation at hand. Perhaps the most 1:34 powerful aspect of agents is their 1:36 ability to reason about paths forward. 1:39 It doesn't just execute a procedure. It 1:41 evaluates possibilities and selects the 1:44 most promising. This isn't just running 1:46 a function. It is reasoning about which 1:49 function to run, when to run them and 1:51 how to interpret their results. So what 1:54 exactly are agentic workflows? It is a 1:56 sequence of operations in which the AI 1:58 drives the process forward. Instead of 2:00 working with if this then that we create 2:03 loops in which the agents are actively 2:06 observing, learning, deciding, and 2:08 acting over and over until it 2:11 accomplishes the goal. But don't confuse 2:13 it with agents. Remember agents are a 2:15 single intelligent function that can 2:17 perceive inputs and decide action. 2:20 Agentic workflows on the other hand are 2:22 orchestrated systems where the agent 2:24 determines which functions to call next 2:26 in what order and with what parameters. 2:29 The AI is actively participating in this 2:31 process, making decisions and adapting. 2:34 It is autonomous and takes actions based 2:36 on context and feedback. These workflows 2:39 are already out there in the wild. For 2:41 example, GitHub copilot is a tool for 2:43 code completion. You start coding. 2:45 Copilot autocompletes based on patterns. 2:48 You accept, reject or modify according 2:51 to the suggestions and co-pilot learns 2:53 from the contact. The agent here is open 2:56 AI codeex and the agentic workflow is 2:59 the collaboration between you and the 3:01 co-pilot. However, how do you determine 3:03 if something is agentic? It needs to 3:06 perceive context meaning understanding 3:09 the code structure. Iterate and improve 3:11 meaning take the user feedback into 3:14 account. Decide the best action. So 3:16 generate code. 3:18 Actutonomously meaning modify code not 3:21 just suggested though. How does it take 3:24 action for each of the steps? Well, in 3:26 order to perceive the AI takes the user 3:28 input and it looks at everything around 3:29 it. So it looks at the files, the code 3:31 snippet and the functions. Then it 3:34 remembers the interactions you had 3:35 within the session in order to give you 3:37 good suggestions and to keep 3:38 consistency. Then it takes the code and 3:41 it actually compares it to similarities 3:43 within the pre-trained knowledge it has 3:45 in order to go down the most probable 3:46 path. From there it does reasoning. So 3:49 AI met for all heristic steps like is a 3:52 loop inside a loop and then it suggests 3:54 some 3:55 optimization. All the while it does 3:58 reinforcement learning at the end. It 4:00 actually gives you a lot of good 4:02 options, but it chooses the one based on 4:04 accuracy, efficiency, and you know, past 4:07 engagements with you. When it comes to 4:09 changing the code, it actually looks at 4:11 what you've written or what you've given 4:13 it. So that way it looks at like the 4:15 cons, what are the function names and 4:18 like what is the syntax itself. It 4:20 compares it to the best practices it has 4:22 as well as the coding style that you've 4:25 given it within the instructions. From 4:27 there it actually looks for you know 4:30 testing of visibility so that way it 4:32 doesn't break any dependencies but at 4:34 the end of all you know it actually 4:36 comes down to the agent whether it wants 4:37 to change that specific section or if it 4:39 want to change the whole thing. The ship 4:41 happening now isn't just technical it is 4:43 a new relationship with code. The future 4:45 isn't replacing code with AI. It is 4:47 understanding where precision matters 4:49 and where adaptability creates value. It 4:51 is all a spectrum and you get to decide 4:54 where each part of your system belongs 4:56 in that spectrum. All in all, the most 4:58 powerful systems won't be done by 5:00 somebody who abandons what they know, 5:02 but by somebody that is willing to 5:04 bridge what was with what is become. 5:09 [Music] 5:10 [Applause]
