WEBVTT 00:00.360 --> 00:05.120 Traditional LM applications operate in a fundamentally limited way. 00:05.520 --> 00:07.400 As shown on page one of the deck. 00:07.760 --> 00:13.280 They respond to a single query, generate an output, and then terminate execution. 00:13.840 --> 00:19.280 This one and done interaction works well for simple question answering or content generation, but it 00:19.320 --> 00:23.120 quickly breaks down when real world complexity is introduced. 00:23.720 --> 00:26.000 Real tasks require far more. 00:26.440 --> 00:31.960 They demand autonomy, the ability to make decisions without constant human input. 00:32.560 --> 00:39.280 They require persistence, maintaining context across multiple interactions, and they require strategic 00:39.280 --> 00:43.720 thinking over extended time frames, not just single responses. 00:44.320 --> 00:51.320 This slide highlights a fundamental architectural shift moving from single turn intelligence to continuous 00:51.480 --> 00:52.800 autonomous systems. 00:53.240 --> 00:59.640 Instead of treating AI as a tool that answers questions, we begin treating it as a system that can 00:59.750 --> 01:00.350 operate. 01:00.750 --> 01:06.150 This evolution represents the next frontier of practical AI deployment. 01:06.670 --> 01:10.030 AI agents are not about making models smarter. 01:10.430 --> 01:16.750 They are about making systems more capable, resilient, and aligned with real world workflows. 01:17.030 --> 01:20.870 Page two defines what an AI agent truly is. 01:21.310 --> 01:24.590 An AI agent is not just a language model with a prompt. 01:25.030 --> 01:30.990 It is a system that perceives its environment, plans actions toward a goal, executes those actions 01:30.990 --> 01:34.870 through tools, observes the outcomes, and adapts over time. 01:35.390 --> 01:37.790 This creates a continuous feedback loop. 01:38.230 --> 01:43.830 The agent learns from successes and failures, refining its behavior instead of starting from scratch 01:43.830 --> 01:44.670 every time. 01:45.270 --> 01:50.390 That ability to adapt is what separates agents from static AI applications. 01:50.870 --> 01:55.150 The slide introduces the fundamental equation that anchors this entire section. 01:55.510 --> 02:00.700 An AI agent equals LM plus tools plus memory plus control loop. 02:01.140 --> 02:02.900 Each component is essential. 02:03.220 --> 02:05.540 Remove memory and the agent cannot learn. 02:05.740 --> 02:07.900 Remove tools and it cannot act. 02:08.140 --> 02:11.460 Remove the control loop and it cannot operate autonomously. 02:11.900 --> 02:18.380 The visual breakdown perceiving, planning, executing, and adapting shows how agents function as complete 02:18.380 --> 02:19.100 systems. 02:19.540 --> 02:21.260 This is not an abstraction. 02:21.420 --> 02:25.460 It is a concrete architectural model that enables real autonomy. 02:25.700 --> 02:31.780 This slide draws a critical architectural distinction that every engineer must understand. 02:32.180 --> 02:38.860 As explained on page three, prompt based systems and agentic systems are not minor variations. 02:39.020 --> 02:41.460 They are fundamentally different paradigms. 02:41.860 --> 02:44.140 Prompt based systems are stateless. 02:44.420 --> 02:48.460 They generate one response and forget everything immediately after. 02:48.860 --> 02:54.660 They have no memory, no long term goals, and no ability to pursue objectives over time. 02:55.140 --> 02:58.540 They react to inputs but cannot act proactively. 02:59.290 --> 03:03.250 Agentic systems, by contrast, are stateful and persistent. 03:03.610 --> 03:10.330 They maintain context across interactions, execute multi-step workflows, and operate with clear goals 03:10.770 --> 03:15.050 instead of simply responding, agents act with purpose and autonomy. 03:15.570 --> 03:22.170 The key difference stated on the slide is crucial agents act while prompt based systems respond. 03:22.530 --> 03:28.290 This distinction determines whether your AI application can handle complex workflows like investigations, 03:28.290 --> 03:34.810 planning, monitoring, or long running processes, or whether it remains limited to simple Q&A. 03:35.490 --> 03:42.570 Page four breaks down the five essential components that every functional AI agent must have. 03:43.210 --> 03:47.370 These are not optional features, they are architectural requirements. 03:47.690 --> 03:53.450 First, the planner determines what actions to take, in what order, and under what conditions. 03:53.970 --> 03:57.730 It breaks down complex goals into executable steps. 03:58.360 --> 04:05.640 Second, memory stores context, conversation, history, and learned experiences, enabling continuity 04:05.640 --> 04:06.480 and learning. 04:07.040 --> 04:13.560 Third, the executor interfaces with tools and external systems to carry out actions in the real world. 04:14.040 --> 04:20.600 Fourth, the observer evaluates outcomes, detects errors, and feeds results back into the system. 04:21.080 --> 04:27.560 Finally, the controller manages the overall loop, coordinates components, and enforces boundaries. 04:28.080 --> 04:31.040 The critical insight at the bottom of the slide is essential. 04:31.400 --> 04:34.440 Agents are complete systems, not just models. 04:34.840 --> 04:41.600 The LM may be the brain, but without these supporting components, true agentic behavior is impossible. 04:41.880 --> 04:48.520 Planning is where an agent's intelligence becomes visible, as described on page five. 04:49.040 --> 04:51.240 Planning answers key questions. 04:51.680 --> 04:56.200 What steps are needed, in what order, and with what dependencies? 04:56.920 --> 04:59.220 There are two major planning styles. 04:59.500 --> 05:04.260 Reactive planning makes decisions step by step as new information arrives. 05:04.660 --> 05:08.100 This is flexible and adaptive, but less predictable. 05:08.620 --> 05:15.220 Deliberative planning creates a full plan upfront, allowing for optimization and conflict resolution 05:15.220 --> 05:17.140 before execution begins. 05:17.940 --> 05:20.740 The slide highlights three core planning techniques. 05:21.180 --> 05:25.500 Task decomposition breaks large goals into manageable subtasks. 05:26.140 --> 05:31.740 Goal refinement transforms vague objectives into specific, actionable targets. 05:32.260 --> 05:36.700 Conditional branching allows agents to adapt their path based on outcomes. 05:37.300 --> 05:42.740 The choice of planning strategy depends on predictability, resource constraints, and environmental 05:42.740 --> 05:43.580 uncertainty. 05:44.340 --> 05:51.500 Strong planning is what allows agents to move from reactive behavior to intentional, goal driven execution. 05:51.700 --> 05:57.010 Memory is what transforms a stateless language model into a persistent learning system. 05:57.530 --> 06:03.330 Page six explains that without memory, an agent would be trapped in the present moment, unable to 06:03.370 --> 06:05.050 build on past experiences. 06:05.650 --> 06:07.930 The slide outlines three types of memory. 06:08.250 --> 06:15.250 Short term memory holds immediate context and conversation state, enabling coherent multi-turn interactions. 06:15.770 --> 06:22.050 Long term memory stores persistent knowledge in vector databases or traditional storage, allowing recall 06:22.050 --> 06:23.130 across sessions. 06:23.730 --> 06:29.570 Episodic memory records past actions and outcomes, enabling learning from experience and avoidance 06:29.570 --> 06:30.890 of repeated mistakes. 06:31.650 --> 06:34.450 The Lie Memory Matters section is critical. 06:34.850 --> 06:38.490 Memory enables continuity, learning and personalization. 06:38.890 --> 06:43.810 Without it, every interaction starts from zero and the system can never improve. 06:44.370 --> 06:50.570 In practice, memory architecture is often the difference between a chatbot and a true digital assistant. 06:51.170 --> 06:58.600 Execution is where planning meets reality, as shown on page seven, agents execute plans by interfacing 06:58.600 --> 07:05.240 with API's calling functions and interacting with external systems such as databases and services. 07:05.840 --> 07:08.760 The slide illustrates a clear execution flow. 07:09.320 --> 07:13.520 The agent decides what action to take based on goals and context. 07:14.080 --> 07:18.080 The system executes that action using controlled infrastructure. 07:18.720 --> 07:22.720 The result is then validated and fed back into the decision loop. 07:23.320 --> 07:25.120 Three properties are emphasized. 07:25.360 --> 07:29.320 Execution must be controlled, observable and validated. 07:29.760 --> 07:35.480 Safety constraints, rate limiting, error handling, and rollback mechanisms all live in this layer. 07:36.080 --> 07:39.920 The critical principle at the bottom of the slide is non-negotiable. 07:40.320 --> 07:45.200 Agents decide what to do, but external systems execute those decisions. 07:45.840 --> 07:51.920 This separation is essential for security, auditability, and reliability in production systems. 07:52.160 --> 08:00.070 The final slide summarizes why AI agents represent a paradigm shift as outlined on page eight? 08:00.390 --> 08:08.230 Agents transform llms from impressive but limited tools into autonomous digital workers capable of handling 08:08.230 --> 08:10.870 complex, multi-step workflows. 08:11.470 --> 08:17.910 Agents operate autonomously, pursue goals intentionally, and maintain state across interactions. 08:18.390 --> 08:24.950 They differ fundamentally from prompt based systems in architecture, capability, and operational model. 08:25.510 --> 08:31.950 Most importantly, the slide reinforces that planning, memory, and execution are non-negotiable. 08:32.270 --> 08:38.310 Remove any one of these components and you no longer have an agent, just a sophisticated chatbot. 08:38.790 --> 08:42.430 The final insight is powerful and should resonate with learners. 08:42.990 --> 08:47.790 AI agents turn llms into digital workers, not chatbots. 08:48.150 --> 08:54.110 This understanding sets the stage for building truly autonomous production grade AI systems.