Claude Code Certified Architect — Foundations (CCA-F)

Build the mental model that separates AI agents from chatbots and workflows. In one lesson: the perceive-decide-act loop, tool schemas + MCP, the ReAct/tool-use loop and tool_choice, the four layers of agent memory, multi-agent orchestration topologies (supervisor / planner-workers / graph), and production evaluation + observability + red-teaming. Beat-format teaching lesson with 6 topics, 48 beats, 12 diagrams, 6 assessments (5 correct required).

Learn how to compose multiple Claude agents into a single system using hub-and-spoke, orchestrator-worker, and pipeline patterns. By the end you'll be able to pick the right orchestration topology, decompose tasks across specialist agents, and coordinate them via message passing, shared state, or events. Includes 5 assessment questions.

Learn to shape and instrument agent loops with lifecycle hooks and session state. By the end you'll wire PreToolUse/PostToolUse/Stop hooks around a tool-use loop, persist and resume sessions, and cap runaway loops with turn budgets and hook-driven guardrails. Includes 5 assessment questions.

Learn how agent systems fail and how to make them graceful under load. By the end you'll classify errors (transient, permanent, policy), design retries with jittered backoff and circuit breakers, and place fallbacks and escalation paths without swallowing signal. Includes 5 assessment questions.

Learn the deployment patterns that turn a working agent into a production service. By the end you'll design queue-backed worker pools, add observability (logs, metrics, traces), plan cost controls, and manage rollout, versioning, and safe fallback strategies. Includes 5 assessment questions.

Learn what Claude Code is, how it composes CLAUDE.md, tools, and MCP servers into a single loop, and how CLAUDE.md scopes cascade (managed → project → local → user). By the end you'll launch Claude Code, author an effective CLAUDE.md, and read /status to understand which sources loaded. Includes 5 assessment questions.

Learn advanced Claude Code configuration: the settings.json precedence chain, permissions.allow/deny, hooks (PreToolUse, PostToolUse, UserPromptSubmit, Stop, etc.), and skills vs. legacy custom commands. By the end you'll author project settings, wire event hooks, and validate the loaded config with /status and claude doctor. Includes 5 assessment questions.

Learn how Claude Code's built-in tools compose into real workflows — Read/Write/Edit for surgical file work, Bash for shell operations, worktrees for parallel branches, and subagents for isolated context. By the end you'll pick the right tool per task and orchestrate multi-step workflows with subagent forks. Includes 5 assessment questions.

Learn how Claude Code deploys inside an enterprise: managed-settings policies, deny lists, RBAC, secret handling, and audit trails. By the end you'll design a policy that ships to a team via managed settings, enforce required deny rules, and reason about what stays local vs. shared. Includes 5 assessment questions.

Learn to author production-grade prompts: system/tool/user roles, few-shot exemplars, prefill for shape control, and stop sequences. By the end you'll build a repeatable prompt template, apply the "SHOW don't TELL" principle, and know when to reach for prefill vs. structured output. Includes 5 assessment questions.

Learn to force Claude's output into strict JSON via tool_use schemas and response_format. By the end you'll design tight input_schema types with enums and constraints, validate outputs against a JSON Schema, and know when to reach for tool_use vs. plain JSON extraction. Includes 5 assessment questions.

Learn advanced prompting techniques: prompt chaining, batch API for parallel work, extended thinking, and prompt caching for cost/latency wins. By the end you'll compose a multi-stage pipeline, cache the stable prefix, and batch parallel classifications correctly. Includes 5 assessment questions.

Learn to test and optimize prompts like production code: golden test sets, regression harnesses, LLM-as-judge, and A/B evaluation. By the end you'll build a prompt eval harness, quantify quality with pass rates, and decide when to iterate on the prompt vs. the model. Includes 5 assessment questions.

Learn the Model Context Protocol: host/client/server architecture, the three primitives (tools, resources, prompts), and the JSON-RPC frames that flow between them. By the end you'll add an MCP server to Claude Code with claude mcp add, inspect it with the MCP Inspector, and read a real tools/list + tools/call trace. Includes 5 assessment questions.

Learn the tool-design craft: precise names, tight input_schema constraints, informative descriptions, error contracts, and annotations (readOnlyHint, destructiveHint). By the end you'll design a tool that Claude reaches for correctly and recovers from cleanly, avoiding common footguns like optional-vs-required and ambiguous enums. Includes 5 assessment questions.

Learn production MCP server patterns: Streamable HTTP transport (SSE is deprecated), OAuth 2.1 + PKCE authorization, multi-server composition, .mcp.json project scope, and rate/scope guardrails. By the end you'll pick the right transport, add authenticated servers with claude mcp add --transport http, and reason about scope precedence. Includes 5 assessment questions.

Learn to reason about the context window: prompt caching, /compact, sliding-window summarization, and extended context on Opus 5. By the end you'll pick a window-management strategy per workload, cache the stable prefix, and know when to compact vs. clear vs. summarize. Includes 5 assessment questions.

Learn to instrument an agent system for production: three pillars of observability (logs, metrics, traces), SLOs and error budgets, structured logging with correlation IDs, and graceful degradation under partial failure. By the end you'll wire OpenTelemetry into an agent, define user-visible SLIs, and design a health-check strategy. Includes 5 assessment questions.