Bounded or augmented call
Control: Your application makes a defined request; retrieval, tools or memory may enrich the call.
Good starting fit: A bounded task where one response can meet a measured quality target.
Claude Certification Blog
When not to build an agent: workflow vs agent architecture in Claude
Do not build an agent because a task sounds complex. Start with the simplest architecture that can meet the requirement, then add model-directed control only when representative evidence shows that fixed paths are insufficient.
Three architecture optionsCurrent July 2026 scopeEvidence before autonomy
When should you not build an agent? When a bounded call or predefined workflow meets the quality, safety and operating requirements with less complexity. Anthropic's guidance is to find the simplest solution possible and increase complexity only when needed. That is a starting principle, not a published five-factor exam algorithm, and not proof that one predictable step automatically rules an agent in or out.
The defensible choice comes from a representative evaluation: define the outcome, test the simplest candidate, inspect failures, and add autonomy only when runtime adaptation fixes failures that matter. Cost, latency, risk and observability remain design constraints throughout that process.
27% CCAR-F Agentic Architecture domain
14.7% CCDV-F Agents and Workflows domain
17% CCAR-P Solution Design domain
0 standalone CCAO-F agent objectives
A bounded call, a workflow and an agent are different options
Anthropic's Building effective agents draws the key boundary: workflows follow predefined code paths, while agents dynamically direct their own processes and tool use. The same article starts one level lower, with an augmented LLM that may have retrieval, tools or memory, and says a well-designed single call is enough for many applications.
“Predefined” does not mean one straight line. Code can route inputs, run independent work in parallel, enforce gates and repeat a bounded loop. Conversely, an agent is not unconstrained: its tools, permissions, budgets, stop conditions and approval requirements still belong to the application. Hybrid systems are normal when one stage needs model-directed exploration and another needs deterministic enforcement.
Make the architecture decision with evidence, not a magic order
The public certification guides do not prescribe five factors in a fixed order or say the first factor decides the mark. The comparison below is Cred Farmer's design aid, synthesized from Anthropic's public architecture guidance and the current exam objectives. Read across each signal; do not treat a single cell as an automatic verdict.
| Decision signal | Bounded call | Workflow | Agent |
|---|---|---|---|
| Can one bounded response meet the requirement? | Start here and evaluate it. | Add only if a repeatable process improves results. | No benefit yet demonstrated. |
| Are the steps and branches substantially knowable? | Good for one bounded transformation or decision. | Strong fit for explicit paths, gates and retries. | Useful only if runtime adaptation adds measured value. |
| Do intermediate findings change what must happen next? | May work for shallow, bounded variation. | Works when the possible routes can be encoded. | Stronger candidate when the path cannot be usefully predefined. |
| Can the system take costly or irreversible action? | Keep action outside the response contract. | Use programmatic checks and approval gates. | Constrain tools, permissions, budgets and escalation. |
| Can you trace, evaluate and stop the behavior? | Log inputs, outputs and grader results. | Observe each branch and transition. | Require tool traces, stop conditions and outcome evals. |
Evaluate all candidates against the same task set and graders. Measure completion quality, unsafe or unauthorized actions, latency, model and tool calls, review effort, retries and recovery behavior. An agent may outperform a workflow enough to justify more calls; a workflow may deliver the same outcome with easier debugging. The label cannot answer that empirical question.
The steps you can name in advance are the whole test
Ask first whether you could list the steps before starting. If you could, the workflow wins and nothing else needs settling: not spend, not tooling, not what your team already knows. Those questions are real, and they belong to a comparison this design never reached. Wanting something adaptable is not an argument on its own; it has to name a case a fixed sequence would get wrong.
Five workflow patterns are examples, not an exam answer key
Anthropic's public article describes prompt chaining, routing, parallelization, orchestrator-workers and evaluator-optimizer. These are composable patterns observed in production, not an exhaustive catalogue and not a rule that every problem must select exactly one. The important distinction is what each pattern controls.
- Prompt chaining fits fixed, cleanly decomposable stages and can place programmatic gates between them.
- Routing sends a classifiable input to a specialized path.
- Parallelization handles independent sections or multiple attempts concurrently before aggregation.
- Orchestrator-workers lets a central model determine subtasks dynamically and synthesize worker results.
- Evaluator-optimizer repeats generation and feedback when criteria are clear and iteration measurably helps.
Notice that orchestrator-workers is a workflow in Anthropic's taxonomy even though a model dynamically determines its subtasks. Architecture labels describe where control lives, so reason about the actual boundary instead of equating “multiple models” with “agent.”
More autonomy needs stronger boundaries, not less structure
A model-directed path can still produce a detailed event and tool trace; it is misleading to say that an agent has no steps or cannot show what approved an action. The real issue is whether your implementation records those decisions and whether policy-critical approval is delegated at all. For consequential or irreversible actions, keep authorization, validation and approval outside model judgment where possible.
Constrain
Give the least tool and data access needed; separate read, draft and execute permissions.
Observe
Capture model turns, tool inputs and results, state changes, stop reasons and approvals.
Stop
Set budgets, loop limits, timeouts, cancellation paths and machine-readable completion states.
Evaluate
Test outcomes and trajectories, including partial failure, missing evidence and unsafe actions.
Use multiple agents for useful decomposition, not just context limits
A single context limit is not the only defensible reason to split work, and how the exams examine decomposition changes with the track, as does the choice itself, since CCAR-P names three patterns rather than two. Once you have decided to split it, the next questions are what each subagent is told and who may talk to whom. Anthropic's multi-agent research system uses parallel agents for breadth-first questions with independent directions, while separate contexts, prompts, tools and exploration paths create useful separation of concerns. The current multiagent documentation likewise describes parallel, context-isolated threads coordinated by another agent.
Those benefits are conditional. Multi-agent work adds coordination, token use, evaluation and reliability challenges. Partition work only where the units are independently valuable or meaningfully specialized; define what each worker receives and returns; track dependencies and dispatched-versus-returned coverage; and evaluate the final synthesis against the requested scope. A coordinator may also research or refine a plan; there is no universal rule that it can never perform a unit itself.
Where architecture selection appears in the current exams
The July 2026 public guides distribute this topic differently. Approximate domain weights describe a collection of objectives; they cannot be converted into a guaranteed number of questions about agents.
| Exam | Current published location | Published weight | Scope boundary |
|---|---|---|---|
| CCAO-F | No standalone agent/workflow objective | Not separately weighted | Do not import architect or developer objectives into this blueprint. |
| CCDV-F | Domain 1: Agents and Workflows | 14.7% | Includes architecture, construction and patterns; the domain weight is approximate. |
| CCAR-F | Domain 1: Agentic Architecture and Orchestration | 27% | Includes loops, dynamic decomposition, subagents and coordinator patterns. |
| CCAR-P | Domain 1: Solution Design and Architecture | 17% | Covers choosing workflow, agentic and augmented-LLM architectures among broader design work. |
Check the current source for your track: CCAO-F, CCDV-F, CCAR-F or CCAR-P. The public guides state selected-response formats and publish objectives and illustrative samples; they do not say candidates lose a mark for failing to name a factor after selecting the correct response.
For the execution contract below this decision, see Claude tool use and structured output. For the evidence layer, continue with how to evaluate Claude output. You can also build a track-specific study plan or compare all four certifications before choosing a track.
Key takeaways
- Start with the simplest sufficient architecture. Prove the need for additional control-flow complexity on representative work.
- Predictable paths favor workflows; changing paths may favor agents. That is a design signal, not a universal disqualifier.
- Measure the trade-off. Compare quality, risk, latency, successful-task cost, review and recovery, not labels.
- Keep consequential controls explicit. Tools, permissions, approvals, budgets, stop conditions and traces belong in the system design.
- Use current track scope. Relevant domain weights are 14.7% on CCDV-F, 27% on CCAR-F and 17% on CCAR-P; CCAO-F has no standalone objective.
Practise architecture decisions against current objectives
Use track-specific study guides and mock exams to test the boundary among bounded calls, workflows, agents and multi-agent systems, then review the objective behind every miss.
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?
Use an agent when model-directed decisions and tool use are necessary because the useful path depends on intermediate findings, and representative evaluations show that the added flexibility beats a simpler design. Predictable work with knowable paths usually fits a workflow better.
Do enumerable steps automatically rule out an agent?
No. Knowable steps are strong evidence for a workflow, not a universal disqualifier. A task may contain both fixed stages and an agentic stage. Choose the simplest architecture that meets the measured requirement and keep high-risk transitions explicit.
Do agents always cost more than workflows?
Not by definition, but agentic systems often trade additional latency and cost for task performance because they may make more model and tool calls. Measure successful-task cost, retries, review and rework on your own workload instead of inferring cost from the label.
What is the difference between a workflow and an agent?
Anthropic distinguishes them by control flow: workflows orchestrate models and tools through predefined code paths, while agents let the model dynamically direct its process and tool use. Both can be simple or complex, and both still need constraints and evaluation.
When do I need more than one agent?
Consider multiple agents when work can be divided into valuable independent directions, specialized roles or context-isolated investigations. Add explicit task contracts, coordination, coverage checks and synthesis evaluation because parallelism also adds cost and failure modes.
Which Claude certification tests agent architecture selection?
The current July 2026 guides place relevant scope in CCDV-F Domain 1 at 14.7%, CCAR-F Domain 1 at 27%, and CCAR-P Domain 1 at 17%. CCAO-F has no standalone agent-architecture objective. Domain weights are approximate, not guaranteed item counts.
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.