0:02 When we build traditional software systems, 0:05 we're architects of certainty. 0:07 Each component, 0:07 each function, 0:09 precise instruction following an exact path, 0:12 reliability through predictability, 0:14 giving the same input we expect the same output 0:18 every 0:18 single 0:19 time 0:21 But, what if 0:22 what if we could explore the best and all path? 0:26 Working with code 0:27 is working, with say, 0:28 Newton's Law! 0:29 Predictable, 0:30 measurable, 0:31 and deterministic. 0:32 Whereas AI components 0:34 follow quantum principles 0:36 probabilistic, 0:37 contextual, 0:38 and sometimes... surprising. 0:44 Let's start with the foundation: 0:46 Large Language Models 0:47 or LLMs. 0:50 These are the engines behind AI systems today. 0:53 Instead of following explicit instructions, 0:55 an LLM predicts what output should come next based 0:59 on patterns in data. 1:01 It is not executing if-else statements 1:03 or for-loops. 1:05 It calculates probabilities 1:06 across BILLIONS of parameters. 1:08 Where our code would crash 1:10 on unexpected inputs, 1:12 LLM gracefully handles ambiguity. 1:14 Where we will write explicit rules for each scenario, 1:17 LLM learns patterns from examples! 1:20 In other words, in traditional software, 1:22 we might build an expert system 1:24 with thousands of explicit rules, hand coded. 1:27 If THIS symptom and THAT symptom, 1:31 then deliver this diagnosis. 1:33 LLM learns relationships from data 1:36 delivering a more nuance 1:38 that no rule system* ever could. 1:41 Now, they're less like 1:42 expert systems 1:44 and more like experts themselves. 1:46 Most of you have worked with
traditional workflows. 1:48 They're the backbone of reliable software. 1:50 Input goes in, 1:52 predictable output comes out. 1:53 In traditional development, 1:55 workflows are deterministic by design. 1:57 They follow explicit rules 1:59 in a predetermined sequence. 2:01 If A, 2:02 then B, 2:02 then C! 2:03 Always. 2:04 Think of your API call, 2:06 database operations, 2:08 or your typical business logic. 2:10 So, we build ETL* pipelines 2:12 to extract data, 2:13 transform it in specific ways, 2:15 and load it where it needs to go. 2:17 Now, what happens when we add an LLM into this workflow? 2:21 This is probably the workflow you most likely interacted with! 2:25 Have you ever had a dreams... 2:27 that... 2:28 This is a workflow that uses AI 2:30 to create content and produce it. 2:32 Now, the key difference is that it's 2:34 probably non-deterministic. 2:36 So that means that the same input 2:38 might produce DifFerEnT 2:39 but still relevant, 2:41 output each time. 2:42 Think of your ETL pipelines 2:44 or your data processing workflows. 2:47 It's predetermined steps 2:48 in a fixed sequence. 2:50 Now... 2:51 GenAI workflows are more like... 2:53 creative production lines. 2:55 The stages might be the same. 2:57 So input, process, output 2:59 but the "process" 3:00 might have an artistic license. 3:06 Taking it one step further, 3:07 if we allowed AI to guide our workflow execution 3:10 based on our inputs, 3:11 those are called Agentic workflows. 3:13 Like GenAI workflows, 3:15 these are also non-deterministic. 3:18 What we are used to are building orchestrated systems 3:21 like Airflow or Prefect 3:23 which coordinate complex sequences of task. 3:26 The orchestrator has a predetermined map they follow 3:29 and if this task fails, 3:31 then try this fallback. 3:33 If it succeeds, 3:34 then try this other thing. 3:35 For agentic workflow, the script is flipped! 3:38 Have you ever had a dreams... 3:39 that... 3:40 that... 3:41 you... 3:42 The AI becomes the orchestrator 3:43 deciding which tools and functions to use 3:46 based on the specific input 3:47 and goals. 3:48 Think of them as less of a... 3:51 big score 3:52 and more of like... 3:53 jazz improvisation! 3:55 The melody is always going to be recognizable 3:58 but the sequence... 4:00 and how it's organized 4:01 is created in the moment! 4:03 Let's be clear about something: 4:04 Agentic workflows are AI orchestrating pre-defined sets of tools 4:09 and functions. 4:10 It selects how they're organized and what sequence 4:14 and which one to use, 4:15 within the boundaries that we've established. 4:17 So again, 4:18 it is about AI having flexibility 4:20 of how they build these workflows specifically 4:23 but, still within the frameworks we have written out for them. 4:29 We finally reach Agents! 4:31 In traditional architecture, we build microservices 4:34 which are independent components that 4:36 handle specific functions. 4:38 Now these communicate through well-defined interfaces 4:40 and are able to follow predetermined logic. 4:43 Agent takes this even further with their autonomy. 4:46 It is a self-contained AI system 4:48 that is able to perceive its environment, 4:50 make decisions, 4:51 and take actions on goals. 4:53 While microservices respond to very specific requests, 4:57 in a... 4:58 you know predictable way 4:59 but the agents 5:01 take this you know even, above and beyond, 5:04 because they're able to 5:05 actively 5:06 pursue 5:06 objectives. 5:08 So they're able to take a tool 5:10 learn from the results 5:11 and then adapt their approach completely, you know. 5:14 Now... 5:14 if I were to make a difference between both of these, 5:17 I would say that the microservices are 5:19 more like a specialized tool that do something really, really perfectly 5:24 and then... 5:25 the Agents are more like a 5:26 craftperson that know 'when' 5:28 and 'how' to use all the tools in their set. 5:31 Have you ever had a dreams... 5:33 that... 5:34 that you... 5:35 um... 5:35 you had... 5:36 Now 5:37 obviously, I have to talk about the difference 5:38 so you don't get confused about Agentic workflows and Agents 5:42 and mainly the difference here would be 5:44 autonomy 5:45 and scope. 5:46 When we talk about Agentic workflows, 5:48 we're talking about the fact that 5:49 it's able to control the whole workflow execution 5:52 with pre-determined tools 5:53 and functions 5:54 to complete a task. 5:56 When it comes to the agent itself, 5:57 it literally just 5:59 tells you "hey, I need a new tool!", 6:01 "I can look up the information by myself", 6:03 "I can select the approach that I want to go about this", 6:06 or, in general, just 6:08 control the whole thing with much greater independence. 6:11 Now, 6:12 obviously, we have to be mindful and think that, 6:14 in the end, when it comes to traditional architecture, 6:17 even all the way down to like Agents, 6:19 we're not talking about replacing one another! 6:22 We're talking about a collaboration that can strengthen them 6:25 because, at the end of the day, 6:27 most of the systems won't be purely deterministic 6:28 or purely AI-driven, 6:31 It'll be a combination of BOTH 6:33 that can be thoughtful and really 6:35 you know holds itself with each other's strength. <3
