Study guide for the largest CCA-F domain: agent loops, subagents, task decomposition, escalation and guardrails, with free sample questions.
Agentic Architecture & Orchestration is the largest block of the Claude Certified Architect – Foundations exam, at 27% of the blueprint. In our practice exam that is 16 of 60 questions. The domain asks a practical question again and again: given a business goal, how should Claude, your code, tools and humans divide the work so the system is useful, safe and affordable?
What the domain covers
The agentic loop. With the Messages API, Claude does not run your tools. It returns a response with stop_reason "tool_use" and one or more tool_use blocks; your application executes them and sends the results back as tool_result blocks in the next user message. The loop continues until Claude ends its turn or your code stops it. Expect questions on what has to be resent (the API is stateless), how to handle several tool calls in one turn, and what a "max_tokens" stop really means — the output was cut off, not finished.
Bounded loops and budgets. Production agents need hard limits: maximum iterations, cost or time budgets, and a defined path when a limit is hit, usually escalation with the work so far. A prompt that says "avoid unnecessary tool calls" is guidance; a counter in the orchestrator is a control. The exam rewards answers that enforce limits in code.
Workflows versus agents. Not every problem needs autonomy. When the steps are known — classify, extract, store — a fixed workflow that calls Claude at each step is cheaper, faster and easier to test. Autonomous agents earn their cost when the path cannot be predicted. Choose the simplest pattern that solves the stated problem.
Coordinator–subagent designs. A coordinator splits broad work into independent subtasks, gives each subagent a clear brief (objective, scope, sources, output format, stopping rule) and merges the results. Subagents keep their own context, so they should return compact, structured findings rather than raw transcripts. Independent subtasks run in parallel; dependent ones run in order. Claims must keep their sources across agent boundaries, and conflicting findings should be surfaced, not averaged away.
Escalation and human-in-the-loop. Know the triggers — an explicit request for a person, repeated failed attempts, policy thresholds such as refunds above a limit, low-confidence or out-of-scope cases — and what a useful handoff contains: a structured summary of who, what was tried, what the customer wants and why the agent escalated, plus the transcript for detail.
Guardrails that do not depend on the model. Authorization, money movement and irreversible actions belong in application code and permission layers. Identity for data lookups comes from the authenticated session, not from what a user types. Agents get the least privilege the task needs, and actions with external side effects can require approval.
How questions are framed
Most questions describe a symptom in a running system — runaway cost, a wrong refund, a coordinator whose context fills up, vague subagent output — and offer four plausible fixes. Usually one fix is a prompt tweak, one is an over-engineered rebuild, one is irrelevant (a bigger max_tokens, a different temperature) and one changes the architecture at the right layer. Look for the answer that makes the failure structurally impossible or recoverable, at proportionate cost.
Study checklist
- Trace a tool-use conversation by hand, including two tool calls in one turn and a tool error.
- List three limits every agent loop should have and what happens when each is reached.
- Write a subagent brief and the structured output contract it returns.
- Decide, for five tasks you know, whether a workflow or an agent fits, and why.
- Map where each business rule in a support agent is enforced: prompt, tool handler, hook or human.
Domain weights and the 60-question form size come from published summaries of Anthropic's exam guide, which is available to registered partners; see our exam guide page for what is confirmed by Anthropic's public FAQ.
Free sample questions: Agentic architecture
Original questions written for practice, not real exam items. Each one appears again, with fresh option order, in the full practice exam.
Sample question 1
Scenario: Customer support resolution agent
A support agent built on the Claude API calls tools to look up orders and issue refunds. In testing, the loop sometimes keeps calling tools for 40+ turns when a customer record is malformed, running up cost. What is the most appropriate first change?
- A. Raise max_tokens so Claude can complete its reasoning in fewer, longer turns per request
- B. Add a system prompt rule telling Claude to stop after a few unsuccessful tool calls in a row, then apologise
- C. Switch to a larger model, which is less likely to repeat the same failing tool call
- D. Enforce a maximum iteration count in the orchestrator and escalate to a human when it is reached
Show answer and explanation
Answer: D. A bounded loop is a basic production control: an iteration (or cost) cap in code makes runaway behaviour impossible whatever the model does, and escalating turns the failure into a handled case. A prompt rule is guidance, not enforcement. max_tokens limits one response, not the number of turns. A bigger model may loop less but guarantees nothing and costs more.
Sample question 2
Scenario: Customer support resolution agent
Your orchestrator receives a response whose stop_reason is "tool_use". The content contains two tool_use blocks. What must the next request to the Messages API contain for the loop to continue correctly?
- A. A tool_result for the first call only; Claude will request the second tool again if needed
- B. One user message holding a tool_result block for each tool_use_id, placed before any text
- C. An assistant message that summarises both tool outputs so Claude can continue from them
- D. Two consecutive user messages, each carrying the tool_result for one of the two tool calls
Show answer and explanation
Answer: B. When Claude requests several tools in one turn, your code runs them and returns every result in the next user message: one tool_result block per tool_use_id, with the tool_result blocks first in the content array. Splitting results across messages or omitting one leaves a tool_use without its result, which the API rejects. Tool output is returned by the user role, not summarised in an assistant turn.
Sample question 3
Scenario: Multi-agent research system
A research system uses one coordinator and several subagents that each search a different source. Subagents return their full raw search transcripts, and the coordinator's context fills up after three subagents. Which change best addresses this?
- A. Have each subagent return a compact, structured summary of findings with their sources
- B. Increase the coordinator's max_tokens so it has more room to hold the incoming transcripts
- C. Run every search inside the coordinator's own conversation so all evidence stays together
- D. Cap each subagent at a single search query so its transcript is always short enough
Show answer and explanation
Answer: A. Context isolation is a key benefit of subagents: each explores in its own window and hands back only what the coordinator needs. Distilled, structured findings with source references keep the coordinator small while preserving traceability. Merging everything into one conversation defeats the purpose; max_tokens controls output length, not how much context is available; one query per subagent cripples research quality.
Sample question 4
Scenario: Multi-agent research system
In a coordinator–subagent research pipeline, three subagents investigate independent sub-questions and none depends on another's result. Latency is the main complaint. What should the coordinator do?
- A. Replace the three subagents with one agent that answers every sub-question in a single turn
- B. Lower the temperature on each subagent so they settle on an answer in fewer reasoning steps
- C. Dispatch the independent subagents in parallel and merge their results when all return
- D. Run the subagents one after another so each can build on the previous subagent's context
Show answer and explanation
Answer: C. Independent work is the textbook case for parallel fan-out: total latency approaches the slowest subagent rather than the sum. Sequential execution only helps when later steps depend on earlier results, which the stem rules out. Collapsing into one agent loses context isolation and focus. Temperature affects randomness, not the number of steps or speed.