Claude conceptsClaude Certified Architect – FoundationsAgentic architecture
CCAR-F agentic architecture: loops and subagents
What CCAR-F asks about the agentic loop, stop_reason, coordinators, subagent context, parallel delegation and error propagation in multi-agent work.
The guide names three loop anti-patterns: parsing natural language to decide when to stop, arbitrary iteration caps as the main stop, and treating assistant text as a completion signal. It also warns that subagents do not share memory between invocations.
01 What does CCAR-F test about agentic loops and subagents?
You trace a loop in your notes and cannot say exactly when it should stop. The guide's answer is stop_reason: continue while it is "tool_use" and terminate on "end_turn".
Beyond the loop, expect coordinator questions: decomposing tasks, choosing subagents, passing context and handling their errors.
The loop stops on stop_reason, not on what the text says.
02 How does the agentic loop work?
Tool results are appended to the conversation history so the model can reason about its next action. A tool_use stop reason means you run the tool and return the result; end_turn means you use the response.
When a tool sequence is critical business logic, such as verifying identity before a refund, enforce it in code, because prompts cannot give deterministic guarantees.
Append every tool result and let stop_reason drive the loop.
03 What does the coordinator own?
A coordinator's refinement loop with subagents
- PlanThe coordinator decomposes the task and picks subagents
- DelegateParallel Task calls, each prompt carrying its context
- WorkEach subagent runs in its own fresh conversation
- ReportOnly each subagent's final message returns
- RefineThe coordinator checks for gaps and re-delegates
The coordinator decomposes the task, delegates, aggregates results and picks the subagents a query needs instead of always running the full pipeline. It spawns parallel subagents by emitting multiple Task tool calls in one response.
Subagents do not inherit parent context, so the coordinator puts the findings they need in the prompt. In the Agent SDK, that prompt string is the only content a subagent gets from the parent, and only its final message returns.
If a subagent needs it, it goes in the prompt.
04 What goes wrong in multi-agent workflows?
Overly narrow decomposition leaves broad topics incompletely covered. A generic error like "search unavailable" hides context; with structured error context, the coordinator can proceed with partial results and note coverage gaps.
Know the neighbors too: subagents report back to the conversation that spawned them, while agent team teammates message each other directly.
Pass structured errors up and keep going with what you have.
Summary
- Continue on tool_use and stop on end_turn.
- Enforce critical tool sequences in code, not prompts.
- Give each subagent its context in the prompt.
- Propagate structured errors with partial results.
Sources and editorial status
Research checked Nov 3, 2026.