Description Transcript
In this chapter, we explore how modern AI systems and developer workflows have evolved. Learn about popular AI workflow patterns like GenAI, Agents, and Agentic workflows, plus what you need to know to start building with them. Whether you're a developer, architect, or just AI-curious, this video will answer how we can coexist in this space.
🔖 Chapters
00:00 - Intro
00:44 - What do LLMs do?
01:46 - What is workflow automation?
02:21 - What are GenAI workflows?
03:04 - What is an agentic workflow?
04:27 - What are agents?
05:37 - Agents vs Agentic
06:11 - Will AI replace us?
#ai #coding #agents
Read more 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
