Agents are one of the bigger blocks on the Generative AI Leader exam, and the questions almost always come down to picking the right architecture for a described scenario. There are three architectural styles you can build with: deterministic, generative, and hybrid. Each one has a defining feature, a clear set of strengths, and a clear set of limitations. Once those are locked in, the four canonical agent examples on the exam (customer service, data, code, security) become a quick mapping exercise.
I think of these three styles as a spectrum. On one end you have full control and zero surprises. On the other end you have natural, flexible conversation. In the middle you get the best of both, paid for in extra design complexity.
The defining feature of a deterministic agent is hard-coded logic. Every decision branch is written in advance, so the agent can only do what it has been explicitly programmed to do.
Strengths:
Limitations:
The defining feature of a generative agent is LLM-driven reasoning. Instead of following predefined rules, the agent uses a language model to reason and respond in natural language.
Strengths:
Limitations:
The defining feature of a hybrid agent is intent-based routing. A router detects the user's goal and switches between scripted rules and generative AI as needed, all while maintaining a single conversation flow. Compliance-critical flows like processing a refund or verifying an identity are handled deterministically. General conversation and open-ended questions are handled generatively.
Strengths:
Limitations:
The decision point on the Generative AI Leader exam usually comes down to one sentence in the scenario. If the scenario emphasizes compliance, auditability, or zero tolerance for unexpected outputs, the answer is deterministic. If it emphasizes natural conversation and flexibility, the answer is generative. If it needs both, the answer is hybrid.
The exam pairs the three architectural styles with four real-world agent types. None of the four is purely one architecture by accident. The right choice always follows from what the agent needs to do.
Handles inbound customer inquiries, answers FAQs, resolves issues, processes requests, and escalates complex cases to human agents. Compliance flows like verifying an account or processing a return need scripted paths, while general questions need to feel conversational. Neither architecture alone can do both, so this agent combines scripted compliance flows with conversational AI. That is hybrid.
Core capabilities:
Accesses, processes, and surfaces data from structured and unstructured sources, generates reports, creates visualizations, and answers data queries. The exam is explicit on one point: a data agent does NOT perform security analysis or threat detection. That belongs to a different agent type entirely.
Core capabilities:
Assists developers across the software development lifecycle: writing new functions, reviewing pull requests, debugging errors, explaining unfamiliar codebases, and generating unit tests. None of that can be reduced to predefined rules. It requires the model to reason, adapt, and generate, which is exactly what generative architecture does best.
Core capabilities:
Ingests security data from multiple sources, correlates events to identify attack patterns, distinguishes genuine threats from false positives, and recommends or automates response actions. Pattern matching for known threat signatures benefits from strict rules. Reasoning about novel or ambiguous threats benefits from generative capability. The design works best when both styles are present.
Core capabilities:
When a Generative AI Leader exam item describes an agent, I read it twice. The first read is for the task: what does the agent actually need to do. The second read is for the constraint words: compliance, auditability, predictability, conversation, open-ended, novel inputs. Those words are how the question tells you which architecture to pick. If both sets of words show up in the same scenario, the answer is almost always hybrid.
My Generative AI Leader course covers types of agents alongside the rest of the foundational material.
