Agentic Architecture & Orchestration is the biggest domain on the CCA-F exam at 27% of total marks. It asks you to reason about how Claude-powered agents are structured, how they coordinate with each other, and how you keep long-running pipelines reliable. Nearly every question in this domain presents a broken or sub-optimal system and asks you to diagnose the problem or choose the better design.
An agentic system built on Claude has at least two layers: an orchestrator that holds the overall goal and delegates subtasks, and one or more subagents that execute those subtasks using tools. The exam tests whether you know which layer is responsible for which decisions.
The orchestrator is responsible for task decomposition — breaking a user goal into a sequence or graph of subtasks that can each be handled by a focused subagent. It also owns the lifecycle of the overall job: deciding when a subtask needs to be retried, when to escalate to a human, and when the goal has been achieved. Subagents, by contrast, are narrowly scoped: they execute one capability well and report structured results back to the orchestrator. Conflating these roles is the source of most architectural anti-patterns the exam will present to you.
Topology questions ask you to choose between a centralised orchestrator that routes all work sequentially versus a parallel fan-out that delegates to multiple subagents concurrently. Parallel fan-out is faster when subtasks are truly independent, but it requires the orchestrator to merge partial results, and any one subagent failure may invalidate the merged output. The exam often presents a scenario where a developer chose parallel fan-out but the subtasks have a hidden dependency, producing incorrect results.
Lifecycle hooks let you inject behaviour at key points in an agent's execution: before a tool call is made, after a tool response is received, when a subagent completes, and when the orchestrator decides to terminate. Common uses include logging, rate-limit enforcement, cost accounting, and safety checks. The exam will present scenarios where a hook is placed at the wrong lifecycle point and ask you to identify the problem.
Scenario: A development team is building a research pipeline. The orchestrator sends the user's research question to three subagents in parallel: one searches the web, one queries an internal knowledge base, and one fetches recent financial data. The orchestrator then merges the three results into a final answer. In testing, the pipeline occasionally returns inconsistent financial figures because the web-search subagent sometimes retrieves a cached page with outdated data, and the financial-data subagent retrieves live data, creating a date mismatch in the merged output.
The correct diagnosis is that the two subtasks fetching financial information are not truly independent: they must share a reference date for the merged output to be coherent. The fix is to pass a canonical reference date from the orchestrator to both subagents before dispatching them in parallel, so that both subagents operate on data from the same point in time. The distractor answer — retrying the web-search subagent until it returns a live result — is wrong because it addresses the symptom rather than the architectural cause.