Practice Exams:

Anthropic CCA-F: Agentic Architecture and Orchestration

AI & Machine Learning

Agentic Architecture and Orchestration is the largest Claude Certified Architect – Foundations domain at 27%. The architect’s job is to choose the least complicated pattern that can complete the task safely, then control authority, state, error handling and cost as autonomy increases. An agent loop is justified only when the next useful step depends on what the system discovers. Predictable work can often remain a bounded sequence, while open-ended work requires explicit tool limits, durable state, approval boundaries, and a testable stop condition.

On this page
  1. Select the architecture from the shape of the problem
  2. Use workflows when the path can be specified
  3. Reserve agent loops for genuinely open-ended work
  4. Treat tools as authority boundaries
  5. Put human approval where consequences justify it
  6. Keep application state separate from active model context
  7. Add multi-agent coordination only for a demonstrated reason
  8. Test architecture through failure, not only happy paths

Select the architecture from the shape of the problem

A Claude application can be a direct model call, a deterministic workflow, a model-assisted router, a bounded agent loop or a coordinated multi-agent system. These are different operating models, not stages of automatic maturity. Each step toward open-ended behavior increases flexibility but also introduces more state, tool calls, failure modes and places where an incorrect assumption can compound. Multi-step designs discussed in Claude Certified Architect – Foundations (CCAR-F) need explicit state transfer between steps and a recovery path when an intermediate tool fails.

Anthropic’s Building Effective Agents guidance recommends starting with a simpler workflow before introducing autonomy. Its Claude Certified Architect – Foundations blueprint gives Agentic Architecture and Orchestration a 27% share, but that weighting does not make more complex agents inherently better. For a support request that requires document retrieval and a deterministic eligibility check, a fixed sequence may be easier to validate than an open-ended agent loop. Reserve discretionary tool choice for tasks with genuine branching uncertainty, and measure successful completion, unnecessary calls, authorization failures and recoverability before increasing autonomy.

Suppose a company must classify incoming customer requests and generate an explanation from a small knowledge base. If the categories are stable and the evidence can be retrieved predictably, a bounded workflow may be sufficient: classify, retrieve, draft, check. The broader Claude Production Engineering coverage considers the operational controls that support these patterns. An open-ended agent that chooses any tool in any order could create additional latency without improving the answer. The useful design question is not whether the system feels autonomous; it is what uncertainty the additional autonomy resolves.

Use workflows when the path can be specified

A workflow keeps control flow in application code while using the model for bounded tasks. It can be sequential, parallel, conditional or iterative within explicit limits. A sequence might extract fields from an invoice, validate them, look up a purchase order and route exceptions for review. Each stage has a known contract and the program determines the next permitted step.

Handoffs should be small and inspectable. A classification stage can return an allowed category, confidence indicator and evidence pointers rather than an entire conversation transcript. A retrieval stage can return document IDs, excerpts and permissions. The drafting stage should never be authorized to change the source records merely because it can describe them. This separation makes failure handling more reliable and creates a clear audit trail.

Use workflows when the path can be specified
Pattern Appropriate when Architecture risk to control
Direct response One bounded reasoning step can solve the request Unclear scope or unsupported factual assertions
Sequential workflow The stages and their dependencies are known Error propagation and ambiguous stage contracts
Parallel work Subtasks are independent and outputs can be merged Duplicate work, contradictory results, higher aggregate cost
Routing Inputs can be classified into distinct policies or specialists Wrong route sending a difficult case to an unsuitable path
Evaluator–optimizer A measurable quality criterion supports revision An expensive or endless revision loop
Agent loop The next useful step cannot be pre-scripted Authority creep, budget overruns and unpredictable termination

A route should fail safely when classification is uncertain. For example, ambiguous customer refunds could be directed to a human rather than to the lowest-cost automated path. A parallel review should reconcile disagreements explicitly instead of silently accepting the last response returned.

Reserve agent loops for genuinely open-ended work

An agent loop observes a goal, chooses a tool or action, receives a result, updates its state and chooses the next step. This can be valuable in exploratory troubleshooting, research or coding, where the precise sequence depends on what is discovered. The benefit comes from adaptive choice, not the mere presence of tool calls.

A safe loop needs a definition of completion that exists outside the model. The host application can require a success artifact, a verified target state, an explicit user decision or a bounded diagnostic result. Without these criteria an agent can continue searching after it has enough information, repeatedly retry a failing call or keep revising an adequate answer. Track action count, elapsed time, rate-limit pressure, tool errors, retries and a cost budget appropriate to the business consequence.

An incident-investigation example illustrates the difference. A workflow may always collect endpoint and identity logs before presenting a summary. An agent may adaptively request additional DNS or cloud control-plane evidence when a novel path appears. That adaptation helps only if the system records what it has already checked, avoids duplicate collection and knows when uncertainty should be escalated rather than hidden with another query.

Treat tools as authority boundaries

Tool access is an application security decision. A model may propose a parameterized action, but the tool service and host application must decide whether the actor is authorized, whether the target is within scope and whether the proposed operation is safe in the current state. A tool’s JSON schema can prevent malformed parameters; it does not prove that the operation is permitted.

Sound Claude agent tool design uses narrow schemas, descriptive definitions, and constrained results. In an orchestration design, tools should also communicate success, failure, partial completion and side effects clearly. A search endpoint that returns an empty collection is different from a search endpoint that failed authentication. An agent needs that distinction to avoid interpreting infrastructure errors as evidence that no data exists.

Use read access by default when the task is investigative. Require stronger authorization for changes to payments, privileges, production deployments or external communications. Where a tool can make an irreversible change, build in a preview or dry-run phase and define who may approve the committed action. Never delegate policy enforcement to persuasive prompt wording.

Put human approval where consequences justify it

Human review is valuable before irreversible, high-impact, costly or externally visible actions. It is not useful if a reviewer receives only an opaque prompt saying “Approve agent action?” The approval surface should show the proposed change, target system, parameters, expected effects, evidence and a way to reject or revise. The same approval should not remain valid indefinitely if the underlying resource has changed.

For example, an agent may propose suspending a compromised service account. The reviewer should see which account, which services depend on it, the evidence of compromise, the scope of the action and a recovery path. After approval, the application should re-check current permissions and target state. Otherwise an approval made from stale data can become unsafe even when the original recommendation was reasonable.

The Claude agent approval guidance explains this as a control boundary. Evaluation should measure not only whether approval exists, but whether high-impact actions are actually blocked when approval is missing, denied, expired or tied to a different target.

Keep application state separate from active model context

Conversation context is not a reliable substitute for a state machine. A production agent needs durable records of workflow stage, prior actions, approvals, retries, artifacts, errors and completion conditions. Store exact state in application storage; pass only the relevant summary and evidence to the model at each step. This reduces ambiguity when a session is resumed and limits the damage from truncated or stale conversational history.

Reliable context management selects useful evidence without carrying irrelevant or stale material between workflow steps. In orchestration, a handoff might record a case ID, retrieval snapshot ID, verified facts, unresolved questions and next authorized step. The downstream agent should not have to infer which stage is complete by rereading fifty chat messages. For a multi-agent design, shared state must also prevent simultaneous actors from overwriting one another’s findings.

A good state record distinguishes observed evidence from model inference and approved action. If a tool’s result says a database is temporarily unavailable, that is not the same as a conclusion that the queried record does not exist. Store those states separately so later steps can retry appropriately or report an evidence limitation.

Add multi-agent coordination only for a demonstrated reason

A multi-agent system may divide long work among specialists, maintain different contexts or parallelize independent analysis. It also creates coordination overhead: a manager must assign tasks, supply constraints, reconcile outputs and check that conflicting findings are resolved. Several agents can agree on the same mistaken premise if they share flawed evidence or the same unclear task definition.

Use a small comparison before choosing multi-agent. Ask whether a single agent with better tools and context performs the work as well. Test whether parallel decomposition genuinely lowers elapsed time after coordination and verification costs are included. Consider whether separate permission domains are valuable: a research agent can collect information, while a privileged deployment agent remains inaccessible until approval. Separation can improve safety when enforced in systems, not just described in prompts.

Observe the whole process: routing choices, handoff contracts, tool invocations, approval decisions, timeouts, retry loops and final artifacts. An answer that looks correct is not sufficient if nobody can explain why the system invoked a sensitive tool or where its evidence came from.

Test architecture through failure, not only happy paths

Before release, simulate unavailable tools, slow responses, wrong classifications, conflicting parallel results, missing approvals and repeated retries. Verify that the application can stop, preserve evidence, return a clear status and resume without duplicating side effects. A request to create a ticket must not create five tickets because the response timed out after the first write.

For the architect-level exam, practice choosing the smallest effective pattern for a new scenario and naming what would make you reject it. Can the path be specified? Are subtasks independent? What resource can the model read or change? How will it know the task is complete? Which outcomes require human judgment? What telemetry is needed to investigate a failure?

These questions make the architecture defensible. The goal is not to celebrate maximum autonomy; it is to design a Claude-based system whose usefulness, permissions, evidence and termination behavior remain understandable under pressure.

Continue learning

Related guides

Tool Design for Claude AgentsDesign bounded tool contracts.Human Approval for Claude AgentsPlace review at consequence boundaries.Context ManagementKeep durable workflow state out of bloated prompts.Claude Production EngineeringOperationalize the architecture.

Related Posts

• Claude Enterprise Operations

• Claude Production Engineering

• Anthropic CCA-F: Choosing the Right Claude Model

• Anthropic CCA-F: Human Approval in Claude Agent Architectures

• Anthropic CCA-F: Context Management and Reliability

• Anthropic CCA-F: Cost Control for Claude Workloads

• Anthropic CCA-F: Designing Claude Applications for Production

• The Vital Role of Machine Learning for Data Scientists

• Microsoft SC-401: Protecting AI Data Beyond Traditional DLP

• Claude Development