7. Agentic Patterns & Task Decomposition

Domain 3 (27%) — Agentic Architecture & Orchestration

The four decomposition patterns — Match the pattern to the task type — this is what the exam tests

Pattern	When to use
Prompt chaining	Fixed, repeating workflows with steps known in advance. Code reviews that always check style → security → docs.
Routing	Classify input first, then dispatch. Different input types need completely different handling.
Orchestrator-workers	Central LLM determines subtasks dynamically. When required steps aren't known in advance.
Dynamic decomposition	Agent generates subtasks incrementally based on findings. Best for investigative tasks where each finding reshapes the plan.

Warning

Applying orchestrator-workers to predictable workflows adds unnecessary overhead. If steps are the same every time, prompt chaining is simpler and cheaper.

Sub-agent context transfer — Summary-and-spawn vs. resume

Approach	Notes
Summary-and-spawn	Summarize key findings → new sub-agent with that summary. Best efficiency. Some information loss but avoids overloading with full history.
Resume session	Risk: if codebase changed since last session, transcript may reference renamed functions or moved files.
Parallel agents	Export findings to a file → create two sessions referencing it. Avoids re-reading the same files in each parallel agent.

Built-in tool selection reference — Claude Code / Agent SDK — which tool for which task

Tool	Use
Grep	Search file contents by pattern
Glob	Find files by name/path pattern
Read	Read a specific file
Edit	Targeted edit via unique string match
Read → Write	Full file replacement (when Edit fails on repetitive content)
Bash	Shell commands, system operations