CCA-F Part 2: Agentic Architecture and Orchestration (27% of the exam)

Part 2 of 7 of Claude Certified Architect — Foundations: a complete self-study tutorial. See all parts.

Domain 1: Agentic Architecture & Orchestration (27%)

This is the largest domain. It tests whether you can choose the right agent shape for a problem and reason about how multiple agents coordinate. The exam does not test framework-specific code; it tests architectural judgement.

Core patterns

Single agent with tools. The default. One Claude call, given a system prompt and a set of tools, runs an internal loop: reason → call a tool → see the result → reason again. Use this for any problem that one specialist could solve given the right tools.

Supervisor / worker (orchestrator–specialist). A coordinator agent receives the request, decides which specialist to invoke, and stitches the responses together. Use it when the problem has clearly distinct subdomains (returns vs. billing vs. accounts) or when each specialist needs a different system prompt, tool set, or model.

Plan-and-execute. The agent first produces a plan (often via Claude Code's plan mode), the user approves it, then a second pass executes. Use it for anything where the cost of a wrong move is high — refactors, schema changes, deployments.

Evaluator-optimiser (multi-pass). A first agent produces a draft; a second agent reviews it against explicit criteria. Use it whenever quality matters more than latency: code review, structured extraction, customer-facing responses.

Subagents. A parent Claude call can spawn child Claude calls that have their own context windows. The child runs to completion and returns a short summary. Use subagents when a sub-task would otherwise blow up the parent's context (e.g., "read these 80 files and tell me which ones implement auth").

When to delegate vs. execute locally

Delegate to a subagent when:

• The sub-task generates lots of tokens you will then discard (file reading, log scanning).
• The sub-task needs a different model (e.g., Haiku for a cheap classification, Opus for a hard one).
• The sub-task is parallelisable (research multiple topics at once).

Execute locally when:

• You will use the full result.
• Latency matters and serial reasoning is fine.
• The work needs the parent's memory / scratchpad.

Escalation patterns

Production agents need to know when not to act. The two common patterns are:

1. Confidence-gated escalation. The agent self-rates its confidence (e.g., on a 1–5 scale); below the threshold it routes to a human or to a higher-tier model.
2. Categorical handoff. The agent classifies the request first; certain categories (refunds over £500, legal questions, compliance) are always handed off regardless of confidence.

The exam frequently asks "the agent resolves X% of category A and Y% of category B — what's the likely cause?" The most common answer: category B needs its own specialist or its own escalation rule, not a smarter shared agent.

Anti-patterns the exam punishes

• One giant prompt that tries to handle every category. Always split by category.
• "Smart" agents that retry indefinitely on failure. Always cap retries and escalate.
• Subagents that pass the entire parent context down. Pass only what they need.
• Coordinators that re-run expensive subagents on partial failure. Cache and reuse.

<a id="domain-2"></a>

← Part 1 of 7: Exam Overview and How to Study · All parts · Part 3 of 7: Claude Code Configuration and Workflows (20%) →