Day 1 — Agentic Architecture & Orchestration
Exam domain 1 (27% of scored content). Build agents that decide and act — implement the agentic loop and orchestrate coordinator–subagent systems.
- The agentic loop: stop_reason control flow, executing tools, and returning results for the next iteration
- Model-driven decision-making vs pre-configured decision trees or fixed tool sequences
- Hub-and-spoke coordinator–subagent orchestration with isolated context; decomposition, delegation, and result aggregation
- Subagent spawning via the Task tool, parallel subagents and explicit context passing, sessions (--resume, fork_session), and hook-based enforcement
Exercises:
- The Agentic Loop
- Coordinator–Subagent Delegation with Task
- Parallel Subagents & Context Passing
- Build & Debug a Multi-Agent Research Pipeline
Exam checkpoint: scenario-based quiz on the agentic loop, coordinator decomposition, subagent context, and enforcement vs prompting.
Day 2 — Tools, MCP & Claude Code
Exam domains 2 (18%) and 3 (20%) — 38% combined. Design tools Claude selects reliably, wire in MCP servers, and configure Claude Code for a real team workflow.
- Tool-interface design and diagnosing misrouting from overlapping descriptions
- Structured MCP errors (isError, errorCategory, isRetryable) and scoping toolsets with tool_choice
- MCP integration — project- vs user-scoped servers, env-var secrets — and built-in tools (Grep, Glob, Read/Write/Edit)
- Claude Code configuration (CLAUDE.md hierarchy, path-specific rules, slash commands, skills) and workflows (plan mode, the Explore subagent, CI/CD)
Exercises:
- Fix Tool Misrouting with Descriptions
- Structured MCP Errors & Scoped tool_choice
- Configure Claude Code for a Team Dev Workflow
- Plan Mode vs Direct Execution
Exam checkpoint: scenario-based quiz on tool descriptions, MCP error/scoping, CLAUDE.md hierarchy, path rules, and plan mode.
Day 3 — Prompt Engineering, Structured Output, Context & Reliability
Exam domains 4 (20%) and 5 (15%) — 35% combined. Engineer precise, schema-compliant output and keep multi-step systems reliable — then sit a full mock exam.
- Precision prompting with explicit categorical criteria and few-shot examples for consistent, actionable output
- Guaranteed schema-compliance via tool_use; nullable fields, extensible enums, and validation–retry–feedback loops
- Batch & multi-pass review — the Message Batches API, SLA planning, and independent review instances
- Context management and reliability — trimming verbose output, persisting case facts, escalation triggers, error propagation, and provenance
Exercises:
- Explicit-Criteria Prompts to Cut False Positives
- Structured Extraction with tool_use + Validation-Retry
- Batch Processing & SLA Planning
- Escalation, Error Propagation & Provenance drills