AI Agents Explained: Models, Workflows, and Decision Systems

1. Introduction to the Course: Hi, I'm Dimple Sangui, and I have over 25 years of experience leading large scale transformation programs. I have led these programs across fortune 500 organizations like Cognizant, HSBC, CAP Gemini, and I have also trained thousands of professional in analytics, AI, and lean six sigma. I have also done leadership training globally. In this class, you will learn how to design effective AI agents. Using practical patterns like prompt chaining, routing, evaluators and orchestrators. These are based on real world agent design principles. This class is for professionals and product leaders, consultants, and AI practitioners who want to go beyond prompts and build structured and reliable AI workflows. You don't need advanced coding, just a basic understanding of how large language models work. Your class project is to design a working AI agent system just as a diagram for a real task. Mapping the workflow, writing prompts, and reflecting on your design choices. I have mentioned the details in the project description section of Skillshare. If you want to understand how agentic AI actually works in practice, this class will give you a clear and usable framework. Let's get started. I will see you in the first class. Oh 2. Building effective LLM: Building effective agents. Over the last year, we have worked with dozens of team building large language models, agents across industries. Consistently, the most successful implementations weren't using complex frameworks or specialized libraries. Instead, they were building with simple composable patterns. What are agents? Agents are defined in several ways. Some customers define agents as fully autonomous system that operates independently over extended periods using various tools to accomplish complex task. Others use the term describing more prescriptive implementation that follows predefined workflows. At anthropic, we categorize all these variations as agentic system, but drawn an important architectural distinction between workflows and agents. Workflows are systems where LLMs and tools are orchestrated through predefined code path. Agents, on the other hand, are systems where LLM dynamically directs their own process, tools, uses maintaining control. When and when not to use agents. When building applications with LLM, we recommend finding the simplest solution possible. Only increasing complexity when needed. This might mean not building agentic system at all. Agentic systems often trade latency, cost of better task performance. And you must consider when this trade off makes sense. When more complexity is warranted, workflows offer predictability, consistency for well defined task, whereas agents are better options when flexibility and model driven decision making are needed at scale. For many applications, however, optimizing a single LLM calls with retrieval in context example is usually enough. There are many frameworks when using Augentic system, easier to implement Lang graph from Lang chain. Amazon's bedrock AI agent framework, revert a drag and drop GIM Workflow Builder, Valm another GUI builder for building testing Complex workflow. These framework makes it easy to get started and make it easy to simplify standard low level tasks, like calling LM, defining, praising, chaining to calls together. However, they often create extra layer of abstraction that only observes the underlying prompt and responses while making them harder to debug. They can make it tempting to add complexity when simpler setup would suffice. 3. Prompt Chaining Workflow: And let's take a closer look to a prompt chaining workflow. Many of you have led complex transformation projects where the risk isn't in the big vision, it's in the execution and ensuring that the discipline is followed. The same happens true. So prompt chaining is not just about passing data from one AI call to another. It's about embedding control gates between each steps. And when you would embed compliance check risk controls and quality audits in business process, the first LLM call creates a draft output. This draft doesn't automatically flow forward. It is validated. If it passes, the workflow continues. If it fails, the system exits or escalates. That is how we prevent the compounding errors. For executives, key insight is chain allows us to treat AI like a managed process, not a black box. Each stage can be tuned whether it is extracting data, applying compliance rules or producing customer facing communication. This design makes AI predictable, auditable, and more acceptable to regulators and boards. To connect it to your world, think of healthcare provider operations. An AI extracts patient information. The gate ensures that the HIPA rules are followed. The second AI prepares the treatment summary. The git validates the medical coding. Only then a discharge node is generated for the patient. Each checkpoint ensures accuracy, compliance, and trust. That is why prompt chaining is a critical foundation. It turns experimentation into enterprise grade reliability. How do we look at prompt chaining workflow? Think of it as an assembly line. Instead of asking AI to do things at once, you break the job into stages. Each stage is handled by a separate AI call, and the output of one stage becomes the input for the next. Notice the gate embitter. That's the quality checkpoint. If the AI output at a stage doesn't pass validation, the process stops or loops back. If it passes, it flows forward. Complex tasks become manageable with prompt chaining workflow. Just like in Sigma, you don't just solve everything in one step, you break it down. Quality control is built in. Instead of trusting one AI call, you validate before moving forward. Error containment. If something fails, you catch it early without polluting the final output. Let's think for an example from claim settlement, step one, AI reads and digitizes the claim document. The gate verifies if all the required fields are present. AI prepares risk scoring for settlement. AI extracts applicant's financial data. The gait checks the compliance rule like missing AIC info. In step three, the AI drafts the loan eligibility assessment. Step one, AI crafts a response. Step two, the gate checks the tone, compliance language, and SLA rules. Step three, AI finalizes the message for customer delivery. Think of it like an airport security. You don't go straight from check in to boarding. You go through gates, baggage checks, security scan, boarding pass check. Each gate ensures that next stage is clean. Prompt chaining is AI's version of staged quality control process. Now that we have seen how the chaining creates reliability, we will move to the next and more advanced workflow, autonomous agent where AI starts to self improve in real environment. Think of it as a relay race. One AI output is checked at the gate and only if it passes goes to the next AI step. If it fails, it exits early. This makes the process safer and more reliable. Motor claim. A customer uploads the accident details in the photographs. The LLM one, does the document check. AI extracts the policy details and the accident description. The gate does the compliance check. Did the AI extract all the required fields if the accident date is within the policy validate? If it passes, it goes to the next step. If it fails, it exits or flags to a manual review. The LLM two call checks the damage estimation. AI drafts the cost estimates using the repair guidelines. LLM three drafts the final claim. It creates the settlement summary for the adjuster. Only valid policy compliant claims move forward, saving time and reducing fraud. Now let's take a customer service example. A VIP customer complains, my loan application is stuck for ten days. LLM one classifies AI identifies the query as a VIP escalations about mortgage. The SLA check happens at the gate. Does it meet escalation of 15 minutes criteria? If it passes, it proceeds. If it fails, it gives an alert to the manager. LLM two, we'll draft the escalation note for level two support. LLM three will create a customer reply, a personalized email to the customer, and the output is the right case are escalated fast, avoiding SLA breaches and customer dissatisfaction. 4. The Routing Workflow time: Now let's understand a routing workflow. The engine behind many augentic AI system. You have an input, you have an LM router, and there are three different LLM calls. So how it works, the input comes in, say, a customer query, a claim form or a loan application. The system just doesn't send it blindly to a one model. Instead, it first goes through an LLM call router. The router decides which AI model is best suited for this task. From this, the request can be routed one of several specialized models. LLM call one might handle structured summarization classification. To might handle reasoning, heavy analysis like fraud detection and risk scoring. The three might handle creative communication tasks like drafting customer response. Finally, the best result is passed out as an output. For you, as transformation leaders, the takeaway isn't the arrow on the charts. It's the business advantage. In healthcare provider service, the router would send medical documentation to a model trained on clinical language while routing billing data to a compliance focused model. In a claims processing motor accident, photos will be sent to a vision model for damage estimation, while the policy text would be going to a language model for the eligibility check. In a customer service, quick FAQs would be routed to lightweight model for speed, while sensitive VIP escalation will go to a model which is tuned for empathy. This workflow matters because it prevents a one size fits all approach. Instead, it ensures that the right model is used for the right job, just like the way you wouldn't assign every task in your organization to the same department. Routing is a governance layer for AI workflow. It makes them efficient, accurate and business ready. Now that we understand routing, we want to see how multiple AI agents can work together, coordinate across different steps in a process instead of acting in isolation. Think of it like a traffic controller. The system decides which lane is best for the incoming request instead treating every request the same way. Claim traging a customer Pfizer Motors insurance claim. The LLM router decides which specialized AI model could handle it. If it's a fraud risk claim, the router, it will route it to fraud detection model. If it's a standard low value claim, route it to automation instead of settlement. If it's a complex medical claim, routed to the compliance model, each claim takes the correct path, reducing the manual sorting and errors. I lost my credit card. The LLM router routes it to the right AI, security AI to block the card immediately. FAI explains how to order a replacement. An escalation AI connects you to the fraud desk if unusual activity is detected. Customer gets the right resolution instead of bouncing between the agents. Book me an appointment and send me the lab results. The LLM router will split the task. Scheduling AI, books the doctor's appointment. WICOdEI retrieves the lab results, and billing AI checks the insurance eligibility and coverage. Patient gets all the tasks completed through one interaction routed through the right AI each time. Routing is about efficiency and precision instead of forcing one system to do everything. It routes the right workflow to the right AI just like the lean process does. Reduces rework, misclassification, and speeds up the resolution. 5. The Parallelization: Now let's go to a parallelization workflow. In the last model, we saw how routing sends task to the best model. Here we will see a different pattern, parallelization of workflow. Here, instead of choosing one path, the input is sent to multiple models at the same time. Each model contributes a piece to the puzzle. Then the aggregator combines the result into a final output. Think of this like running multiple teams in parallel to solve the same problem quickly and thoroughly. Why does this matter? Because some business problem benefits from speed and diversity of answers. We understood that paralyzation is running multiple LLM calls, and then the aggregator summarizes it. So let's say the patient has fever, cough, and recent travel history and allergic to antibiotics. The LLM one retrieves patient EHR for past conditions. LLM two checks the guideline for infectious diseases. LLM three flags drugs allergy risk. The aggregator then merges into a safe treatment suggestion for the physician. The physician gets a single advisory reports not fragmented insight. Parallelization is about speed and comprehension instead of waiting for one system to finish before the next starts, a multiple AI runs in parallel and then converges. It's like running multiple specialist team in parallel during the lean transformation cost cutting. We look at routing when the system decides which model to use. Parallezation when multiple model works at the same time. 6. The Orchestration and Routing2: We look at routing when the system decides which model to use, parleization when multiple model works at the same time. Here we are talking about orchestration. How it the input first goes into orchestra. Think of this like a project manager who is then assigning work across different AI agents. The orchestra decides which task should be done first and which model should do it. Each model like LLM one, two, three, tackles different parts of the job. But instead of working in isolation, they are coordinated like cross functional team in your transformation. Finally, a synthesizer pulls everything back together into a single cohent output. The business value is powerful. Patient notes should be split into clinical interpretation, billing code, mapping, and compliance check then synthesized into one complete discharge report. Let's think about motor claim. One model reviews the document, the other checks the fraud. The third draft the customer communication. This synthesizer then combines them into ready to use claim decision package. Et's think about mortgage business. You want income verification, risk scoring, regulatory compliance checks happen in sequence, and orchestrated and synthesized into one loan approval summary. A press release drop might be fact checked by one model adapted at different markets by another and style by social channel of the third. Then synthesized into multi channel package. Orchestration is about coordinating. It's not about computation. It mirrors the way transformation leaders already run across functional programs. 7. Thank you for your time: Before we wrap up, I want to share a little more about myself and how you can stay connected. I'm looking forward to sharing my two decades of experience with all of you. And the experience where I have helped leaders and teams navigate large scale business, digital and artificial transformation project across India, USA, UK, Australia, and Middle East. Along the way, I have trained thousands of professionals in data analytics, lean Sig Sigma, artificial intelligence, prompt engineering, and change management, leadership training, both as corporate programs and open learning communities. If you're watching this as an individual learner, I would love for you to stay connected. You will find links to join my Whatsapp group or telegram community where I would be continuously posting about different opportunities. I regularly share some practical insights, learning resources, and updates on AI analytics and transformation. I also share some case studies on leadership. If you are here as a manager, a consultant, or a learning and development leader, and you are interested in custom design, corporate training on agentic AI, prompt engineering, data analytics, or transformation program, please feel free to reach out to me. I work closely with organizations to design programs that are custom made, that are practical, contextual and business focused. I'm leaving the link to Linden in the discussion section. Thank you for learning with me. I hope this class helps you design an AI system more thoughtfully, and I look forward to staying connected with you. Thank you for your time. Happy learning.