Workflows, agents, and the pattern vocabulary

Every architecture question on this exam is a vocabulary question in disguise. They describe a system, and you have to name it. So learn the names properly, because “it’s an agent” is wrong roughly half the time.

The line that splits everything

Anthropic’s own framing, and it is a clean one:

• A workflow orchestrates models and tools through predefined code paths. You wrote the steps.
• An agent lets the model dynamically direct its own process and tool usage. The model decides the next step at runtime.

The dividing line is who decides what happens next, not how complex the diagram looks. A monstrous 40-node pipeline with hardcoded branches is still a workflow. A twelve-line while loop where the model picks tools until it stops is an agent.

Underneath both sits the augmented LLM: one model call with retrieval, tools and memory attached. That is the base unit. Everything else is composition.

The five compositions worth memorising

Prompt chaining. Split the task into fixed steps, each call feeding the next. Add a programmatic gate between steps if you want to bail early. Best when the decomposition is genuinely known up front.

Routing. Classify the input, then dispatch it to a specialised path. Wins when different input types need genuinely different handling and lumping them into one prompt makes all of them mediocre.

Parallelization. Two flavours and they get asked separately. Sectioning splits a task into independent subtasks run at the same time. Voting runs the same task several times and aggregates, which is what you want for a judgement call where one sample is noisy.

Orchestrator-workers. A model decomposes the task into subtasks it invents at runtime, delegates them, then synthesises. That runtime invention is the difference from routing, and it is what makes this one agentic.

Evaluator-optimizer. One call produces, another critiques, loop until the critic is happy. Needs a critic that can actually tell good from bad, otherwise you have built an expensive way to churn.

The framework question, and where your other study pays

The CCDV-F exam guide, under Agent Patterns and Frameworks, names agentic abstraction frameworks with the examples Strands, LangGraph and PydanticAI.

Sit with that for a second, because it is the single most useful sentence in this course for your particular situation. LangGraph is named, by Anthropic, on an Anthropic certification. Not as a competitor, not as a footnote: as the worked example of what an agentic abstraction framework is.

So the LCAE study is not a parallel track you are running at your own expense. State, nodes, edges, the checkpointer, interrupt() and the resume: that mental model is on this syllabus too, and it is on the syllabus in the exact place where a candidate who has only ever written raw Messages API calls has nothing to say. You are not learning two things. You are learning one thing twice, from two vendors, and the second pass is the one that makes it stick.

What this section actually grades is not framework trivia. It wants you to know what an abstraction framework buys you: a state model, explicit control flow, persistence and resumability, and observability, in exchange for a dependency and a layer of indirection between you and the raw message list. That list should read as a description of things you have already implemented by hand.

And it wants the counterweight, which is just as gradeable: plenty of production agents are a while loop over the Client SDK. Reaching for a graph framework is a decision with costs, not a maturity level. Saying “a while loop is the right call here” is a senior answer, not a naive one.

The judgement they grade

Reach for the simplest thing that works: single call, then augmented call, then workflow, then agent. Agents trade determinism, latency and token spend for flexibility. If you already know the steps, encoding them is cheaper than asking a model to rediscover them on every request.