Claude Certification Blog
When not to build an agent: workflow vs agent architecture in Claude
Workflow vs agent architecture is examined as an elimination, not a preference. The exam wants to know which factor ruled the simpler design out — and if you cannot name one, the simpler design was the answer.
Do not build an agent when the steps can be listed in advance. That test comes first and it is disqualifying, not advisory — before cost, before latency, before which design is more interesting to build. The exam's habit is to grant the least autonomy and the fewest moving parts that still clear the quality bar, and to escalate only when measurement shows the lighter option falling short.
The test that disqualifies an agent
Enumerability. If you can write down the path through the work ahead of time, an agent is ruled out by default, and no amount of "but it might need to adapt later" reopens it. The study guides put the general rule plainly: prefer the design that grants the least autonomy and uses the fewest primitives that still clear the quality bar.
The tell in an exam item is a task described as open-ended that turns out, on reading, to have a small number of knowable paths. Not yet having worked out the steps is a different condition from the steps being undeterminable, and only the second one justifies handing control to the model.
Workflow vs agent architecture is a three-way choice
Most candidates reduce it to two and lose the marks there. The difference between a workflow and an agent is not size, sophistication or how many tools are involved — it is who owns the control flow. In a workflow the path is written in code ahead of time. In an agent the model directs its own path and tool use at runtime.
The third option is the one people skip: a single bounded call, with the right context supplied and no branching control flow at all. It is the cheapest thing to build, test and reason about, and the honest answer to "should this be an agent?" is often that it should not be a workflow either. Five workflow patterns are named in the public engineering guidance the exams draw on — prompt chaining, routing, parallelisation, orchestrator-workers and evaluator-optimiser — and the point of knowing them is to recognise which shape a problem already has, not to have a catalogue to pick a favourite from.
The factor that decides first
Pattern selection is sequential, not a weighing exercise. Five factors get walked in a fixed order, and the first one that rules a pattern out decides the answer.
That is why the scored skill is the elimination rather than the verdict. An answer that arrives at "agent" because the task felt flexible has not done the work, even when "agent" happens to be right.
Two traps for people who already know the patterns
The first is choosing an agent for flexibility alone. Flexibility is an aspiration, not a factor — it names nothing that disqualified the simpler design, which is precisely what the item is asking you to name.
The second is counter-intuitive enough that arguing the obvious way gets you the exam's own wrong answer: agents do not automatically cost more than workflows. Design drives the bill, not the pattern label. A workflow that retries everything on every branch can cost more than a tightly bounded agent, so the argument has to be about what the design does rather than what it is called.
The real cost of an agent is what you can see afterwards
When the model owns the path, the record of what happened reads like a transcript rather than a step log. That is fine until somebody asks which step approved something — and the honest answer is that no step did, because there were no steps.
When one agent genuinely is not enough
The ladder does not stop at "agent", and neither does the discipline. Multi-agent gets its own gate: confirm the work genuinely exceeds what a single context can hold before splitting it, because splitting introduces a failure shape that a single agent does not have.
The shaded row is the one people leave out. An orchestrator that dispatches four units and receives two will happily synthesise a confident answer from two, because nothing in the loop compares what came back against what went out. Nothing errors. The output simply covers less than it claims to, and it reads exactly as well — which is why checking the output is a separate discipline.
Where this is examined
Not a niche topic. Solution design is the largest single domain block on any of the four exams, and agents get a domain of their own on Architect – Foundations.
Weights are approximate and no exact per-form item count is guaranteed — confirm against the official exam guide. The mechanics underneath all of this, the contract by which a model asks for work and your code does it, are covered in Claude tool use and structured output. For how these exams differ, Claude Certified Architect – Foundations and Architect – Professional publish their blueprints, the guide to choosing a Claude certification compares all four, and the Claude certification study guide covers building practice around a domain this heavy.
Key takeaways
- Enumerable steps disqualify an agent. It is a gate, not a consideration, and it comes before cost or latency.
- The choice is three-way. A single bounded call is the option people forget, and it is the cheapest thing to build and inspect.
- Name the factor that eliminated the alternative. The right verdict without the elimination still loses the mark.
- "Agents cost more" is a trap, not a truth. Design drives the bill; the pattern label does not.
- Multi-agent needs a coverage check. Without one, an orchestrator synthesises confidently from fewer results than it sent out.
Practise the elimination, not the vocabulary
Timed mock exams scored per domain, each item explaining why the wrong options fail — including the architecture scenarios, where the plausible answer and the defensible one are frequently the same pattern reached for two different ways.
See the CCAR-P blueprintQuestions
Frequently asked
The follow-up questions people search next.
When should I use an AI agent instead of a workflow?
When the steps genuinely cannot be listed in advance. If you can enumerate the path, an agent is ruled out before cost or elegance enters the discussion — and the exam treats that as disqualifying rather than as one consideration among several.
Is my task actually open-ended, or am I just unsure how to structure it?
A useful test: not yet knowing the steps is not the same as the steps being undeterminable. Tasks called open-ended often turn out to have a handful of real paths once someone looks, and a design built for flexibility you never needed is harder to inspect afterwards.
Do agents always cost more than a workflow?
No, and assuming so is a named trap. Design drives the bill, not the pattern label — a badly-shaped workflow that retries everything can cost more than a well-bounded agent. Argue from what the design does, not from what it is called.
What is the difference between an agentic and a workflow pattern?
Who owns the control flow. In a workflow the path is written in code ahead of time; in an agent the model directs its own path and tool use at runtime. It is not a difference of size, sophistication or how many tools are involved.
When do I need more than one agent?
Only once you have confirmed the work genuinely exceeds what one context can hold. Multi-agent adds its own failure shape — an orchestrator can synthesise a confident answer from fewer results than it dispatched unless something explicitly checks coverage.
Which Claude certification tests architecture pattern selection?
Architect – Professional most heavily: Claude Platform and Solution Design is its largest domain at roughly 32% of the blueprint. Architect – Foundations covers agents in a domain of its own at roughly 13%. Weights are approximate.
Keep reading
Related posts
Not affiliated with, or endorsed by, Anthropic or Pearson VUE. Details are summarised from publicly published program information and can change — always confirm against the official exam guide before booking.