A structured 12-week study program covering all 5 exam domains. Track your progress, study the anti-patterns, and practice with real exam scenarios.
Designing agentic loops, coordinator-subagent (hub-and-spoke) patterns, session state management, task decomposition strategies, context isolation between agents, forking, and workflow enforcement. The heaviest domain — most architectural in nature.
Common Trap
Using few-shot examples to enforce tool ordering — this is a programmatic prerequisite problem, not a demonstration problem.
Writing effective tool interface descriptions, implementing structured error responses for coordinator recovery, scoping tools to specific agent roles, configuring MCP servers, and distributing tool access across multi-agent systems.
Common Trap
Returning empty results on subagent failure — this silently suppresses errors and prevents recovery. Always return structured error context.
CLAUDE.md hierarchies and rule propagation across nested directories, custom slash commands and skill creation, path-specific rules, plan mode vs. direct execution mode, and CI/CD pipeline integration for automated workflows.
Common Trap
Assuming a larger context window fixes attention dilution — it doesn't. The fix is splitting work into focused per-file passes.
Enforcing prompts, not just writing them. JSON schema design that prevents hallucinations, few-shot technique selection, validation and retry loop architecture, output contract design, and making structured output reliable at scale.
Common Trap
Using self-reported confidence scores for escalation routing — LLM confidence is poorly calibrated. The model is most confidently wrong on the hardest cases.
Preserving critical information across long interactions, designing escalation patterns that fire at the right threshold, managing error propagation in multi-agent systems, and handling uncertainty with confidence calibration.
Common Trap
Routing all workflows to the Batch API for cost savings — Batch has no SLA. Blocking workflows require real-time API regardless of cost preferences.