Curriculum
CCAF logo

Domain 1 · 27% of the exam

Agentic Architecture & Orchestration

The largest domain. It covers how an autonomous agent runs, how multiple agents coordinate, how work is decomposed and handed off, how hooks enforce behaviour, and how sessions persist and branch.

  1. 1.1 Agentic Loops

    Drive an agent's run from the API's stop_reason field — continue on tool_use, stop on end_turn — instead of parsing the model's text or capping iterations.

  2. 1.2 Multi-Agent Orchestration

    A hub-and-spoke coordinator owns decomposition, delegation, routing, and aggregation; when coverage is incomplete, the fix is the coordinator's scope, not the subagents.

  3. 1.3 Subagent Invocation & Context Passing

    Subagents run with fresh, isolated context — everything they need must be passed explicitly in the Task prompt; nothing is inherited.

  4. 1.4 Workflow Enforcement & Handoff

    When step ordering has financial or compliance stakes, enforce it with a programmatic prerequisite gate — prompt instructions have a non-zero failure rate.

  5. 1.5 Agent SDK Hooks

    PostToolUse hooks normalize heterogeneous tool results before the model sees them; PreToolUse hooks intercept and block policy-violating calls — both give deterministic guarantees prompts cannot.

  6. 1.6 Task Decomposition Strategies

    Use fixed prompt chaining when the structure is known up front; use dynamic decomposition when subtasks only emerge from what you discover during the work.

  7. 1.7 Session State & Resumption

    Resume when the same thread's context is still valid, fork to explore divergent approaches from a shared baseline, and start fresh with a summary when prior tool results are stale.