course contents
Architect Foundations
sources and errata
Primary sources only — Anthropic's own engineering writing and product documentation, plus the exam guide that defines what the exam covers. No third‑party paraphrase.
Primary sources
- Anthropic — "Building Effective Agents"
Workflows "orchestrated through predefined code paths" versus agents, stopping conditions such as a maximum number of iterations, and prompt chaining with programmatic gate checks between steps. For 5.2: agents that "pause for human feedback at checkpoints or when encountering blockers", and support success measured by resolutions (checked 2026-10-01). Used in 1.1, 1.6 and 5.2. Checked 2026-09-30. - Anthropic Engineering — "How we built our multi‑agent research system"
Subagent scaling figures, the semiconductor duplication case and four-part subtask brief, the lead agent deciding "whether more research is needed", and structured attribution across handoffs. The failure-adaptation and resume-from-error passages and the CitationAgent. Used in 1.2, 1.3, 1.6, 5.3 and 5.6. Checked 2026-09-30. - Claude Docs — "Subagents in the SDK"
The what-subagents-inherit table, the prompt string as "the only content you pass from parent to subagent",AgentDefinitionrequired and optional fields, theAgent/Tasknaming note, and background execution by default. For 5.3: "only its final message returns to the parent" (checked 2026-10-01). For 2.3: an omittedtoolslist inheriting every tool available to subagents, anddisallowedToolswithmcp__server__*patterns (checked 2026-10-01). Used in 1.2, 1.3, 2.3 and 5.3. Checked 2026-09-30. - Claude Docs — "Configure permissions" (Agent SDK)
allowed_toolsauto-approves listed tools, any tool not listed "is still available to Claude", andAgentis among the tools that don't ask before running. Used in 1.2 and 1.3. Checked 2026-09-30. - Claude Docs — "Intercept and control agent behavior with hooks" (Agent SDK)
Callbacks in thehooksoption,HookMatcherwithmatcher,hooksandtimeout, themcp__<server>__<action>matcher form,permissionDecision: "deny"withpermissionDecisionReason,updatedToolOutput, and a timed-outPreToolUsecallback blocking the call. Used in 1.4 and 1.5. Checked 2026-09-30. - Claude Docs — "Hooks reference"
PreToolUseandPostToolUsedefinitions,tool_responsein thePostToolUseinput, and the timeout contrast: settingscommand,httpandmcp_toolhooks let the call continue, while an Agent SDK callback blocks it. Used in 1.4 and 1.5. Checked 2026-09-30. - Claude Docs — "Manage sessions"
What--resumerestores and which launch flags it does not,--continue, session naming withclaude -nand/rename, and--fork-session. For 5.4:/compact [instructions]replaces history with a summary, optionally focused (checked 2026-10-01). Used in 1.7 and 5.4. Checked 2026-09-30. - Claude Docs — "Work with sessions" (Agent SDK)
What a session stores,resumeby session ID,fork_session=True/forkSession: truegiving the fork its own ID, and "Sessions persist the conversation, not the filesystem". For 4.6: resuming restores the full history and a fork starts from a copy of it, so neither is an independent reviewer. Used in 1.3, 1.7 and 4.6. Checked 2026-10-01. - Claude Platform Docs — "Stop reasons and fallback"
All sevenstop_reasonvalues with the docs' one-line definitions, sending the content back onpause_turn, and raisingmax_tokensor continuing the response. Used in 1.1. Checked 2026-10-01. - Anthropic Engineering — "Writing effective tools for agents"
"When tools overlap in function or have a vague purpose, agents can get confused about which ones to use", consolidated workflow tools such asget_customer_context(2.1), and error responses that give specific, actionable guidance instead of opaque codes (2.2). Used in 2.1 and 2.2. Checked 2026-10-01. - Claude Platform Docs — "Handle tool calls"
tool_resultfields (tool_use_id,content,is_error) and the"role": "user"message that carries them, which for a client tool means a new request (4.5). For 2.2:tool_resultcarries only those three fields, and error text should say what to try instead of "failed". Used in 1.1, 2.2 and 4.5. 2.2 use checked 2026-09-30. - Claude Platform Docs — "Parallel tool use"
Onetool_resultpertool_useblock, all together in the next user message. Used in 1.1. Checked 2026-09-30. - Model Context Protocol — "Tools" specification (2026‑07‑28, current)
"Error Handling": protocol errors for unknown tools, malformed requests and server errors;isError: truefor API failures, input validation errors and business logic errors; clients SHOULD pass tool execution errors to the model and MAY pass protocol errors. "Structured Content": a serialized text copy ofstructuredContent. Used in 2.2. Checked 2026-09-30. - Model Context Protocol — 2025‑11‑25 changelog
Input validation errors returned as tool execution errors, not protocol errors. Used in 2.2. Checked 2026-09-30. - Model Context Protocol — "Resources" specification (2026‑07‑28, current)
Resources as application-driven context, listed withresources/list(URI, name, description) and read withresources/read. Used in 2.4. Checked 2026-10-01. - Claude Docs — "Connect Claude Code to tools via MCP"
The local, project and user scope table, with local as the default and whole-entry precedence local > project > user;${VAR}and${VAR:-default}expansion incommand,args,env,urlandheaders, and the literal text sent when a variable is unset; the GitHub server URL and bearer header; tool search on by default; and resources attached with@server:protocol://path. Used in 2.4. Checked 2026-10-01. - Claude Docs — "MCP" (Agent SDK)
.mcp.jsonis loaded when theprojectsetting source is enabled, which it is for defaultquery()options. Used in 2.4. Checked 2026-10-01. - Claude Docs — "Tools reference"
Glob and Grep absent from the default tool set on macOS, Linux and WSL, and how to restore them; Grep on ripgrep, skipping gitignored files; Glob matching paths, sorted by modification time; Edit's read-before-edit, exact and uniqueold_string, andreplace_all; Write overwriting whole files; Read'sPARTIAL view; and Bashcdversusexportpersistence. Used in 2.5. Checked 2026-10-01. - Claude Docs — "How Claude remembers your project"
The CLAUDE.md scope table and root-down load order, subdirectory files that load when Claude reads there,@pathimports (four hops, expanded at launch, no context saving), the 200-line target,.claude/rules/with thepathsglob table and rules that trigger on read, and/memoryversus/context. Used in 3.1 and 3.3. Checked 2026-10-01. - Claude Docs — "Extend Claude with skills"
The project and personal skill directories, the commands-merged-into-skills statement, theSKILL.mdfrontmatter table (context: fork,agent,allowed-toolspre-approving rather than restricting,disallowed-tools,argument-hintas an autocomplete hint,paths),$ARGUMENTS, forked skills running in the background, and a personal skill winning a name conflict. Used in 3.2 and 3.3. Checked 2026-10-01. - Claude Docs — "Choose a permission mode"
Plan mode reading files and running shell commands to explore without editing source, entry viaShift+Tab,/planor--permission-mode plan, and approval exiting plan mode into an editing mode. Used in 3.4. Checked 2026-10-01. - Claude Docs — "Create custom subagents"
The built-in read-onlyExploreandPlanagents, what Explore skips (CLAUDE.md, git status),Plan's role in plan mode, and a general-purpose subagent having every tool (3.2, 3.4, checked 2026-10-01). For 1.3: the Task tool renamedAgentin 2.1.63 withTask(...)kept as an alias, fork subagents, and background subagents. Used in 1.3, 3.2, 3.4 and 5.4. - Claude Docs — "Best practices for Claude Code"
Skills for on-demand workflows versus CLAUDE.md for what applies broadly, explore-plan-implement and "If you could describe the diff in one sentence, skip the plan", giving Claude a way to verify its work, the interview pattern withAskUserQuestion,/clearafter two failed corrections,claude -pin CI, fresh-context review (the Writer/Reviewer pattern), and "Fan out across files": test on a few files, refine the prompt, then run on all. Used in 3.2, 3.4, 3.5, 3.6, 4.5 and 4.6. D3 uses checked 2026-10-01. - Claude Docs — "Run Claude Code programmatically"
-p, exit code 0 on success and non-zero when the run fails, the three--output-formatvalues,--json-schemaoutput instructured_outputread withjq, and what--bareskips, that it is recommended for CI, and that it "will become the default for-pin a future release". Used in 3.6. Checked 2026-10-01. - Claude Docs — "CLI reference"
--print,--output-formatand--json-schemain print mode,--bare, and--append-system-prompt-file. Used in 3.6. Checked 2026-10-01. - Claude Docs — "Claude Code GitHub Actions"
CLI flags passed throughclaude_args, review criteria and project rules defined in CLAUDE.md, and inline PR comments from the review workflow. Used in 3.6. Checked 2026-10-01. - Claude Platform Docs — "Prompting best practices"
Clear and direct instructions and "Claude is smart enough to generalize from the explanation", using examples (relevant, diverse, structured; 3–5, in<example>tags), the self-correction chain as separate API calls, and the self-check line. For 5.1, "Long context prompting": long documents at the top, the query at the end, and source metadata wrapped with each document. Used in 4.1, 4.2, 4.4, 4.6, 5.1 and 5.2. Checked 2026-10-01. - Claude Platform Docs — "Prompting Claude Sonnet 5"
"Code review harnesses": "be conservative" and "only report high-severity issues" are followed faithfully, so precision typically rises while recall can fall, and the advice to state concretely where the bar is. Used in 4.1. Checked 2026-10-01. - Claude Platform Docs — "Prompting Claude Opus 5"
The same code-review advice for Claude Opus 5. Used in 4.1. Checked 2026-10-01. - Claude Platform Docs — "Reduce hallucinations"
"Allow Claude to say 'I don't know'", which anullexample demonstrates, and validating critical output. Used in 4.2 and 4.4. Checked 2026-10-01. - Claude Platform Docs — "Structured outputs"
JSON outputs viaoutput_config.format; compliance "in most cases", with the "Invalid outputs" cases (stop_reason"refusal"and"max_tokens") and enum capitalisation; supported and unsupported schema features; "No retries needed for schema violations"; andclient.messages.parse()with a Pydantic model. Used in 4.3 and 4.4. Checked 2026-10-01. - Claude Platform Docs — "Strict tool use"
strict: trueon the tool definition as grammar-constrained sampling, and the"2"-instead-of-2failure it prevents. Used in 4.3. Checked 2026-10-01. - Claude Platform Docs — "Define tools", "Forcing tool use"
The fourtool_choiceoptions and their defaults, the prefilled reply underanyand a forced tool, and the models and settings on which those two return an error, with the docs' substitutes (2.3, 4.3). For 2.1: the tool-use system prompt built from tool definitions, descriptions as "by far the most important factor in tool performance", and consolidating related operations with anactionparameter. Used in 2.1, 2.3 and 4.3. Checked 2026-10-01. - Claude Platform Docs — "Build a tool-using agent"
Sending an error back in atool_resultwithis_error: trueso Claude can retry with corrected input. Used in 4.4. Checked 2026-10-01. - Claude Platform Docs — "Batch processing"
The 50% price, the 24-hour window and expiry, 29-day results,custom_idmatching in any order, the four result types and which are billed, fixing aninvalid_request_errorbefore resending, asynchronousparamsvalidation, and the "What can be batched" list. Used in 4.5. Checked 2026-10-01. - Claude Docs — "Code Review"
REVIEW.mdskip rules for paths and finding categories, a nit cap, and redefining what Important means. Used in 4.1. Checked 2026-10-01. - Claude Platform Docs — "Using the Messages API"
The Messages API is stateless, so the full conversation history is sent on every request. Used in 4.4 and 5.1. - Claude Platform Docs — "Context windows"
Every message inmessages, tool results included, counts toward the window, and context rot: accuracy and recall degrade as the token count grows. Used in 5.1. Checked 2026-10-01. - Claude Platform Docs — "Context editing"
Theclear_tool_uses_20250919strategy clears the oldest tool results and leaves placeholder text in their place. Used in 5.1. Checked 2026-10-01. - Claude Platform Docs — "Compaction overview"
Server-side compaction replaces the older turns of a conversation with a summary Claude writes on the server. Used in 5.1. Checked 2026-10-01. - Anthropic Engineering — "Effective context engineering for AI agents"
The attention budget and context rot, structured note-taking outside the context window, and subagents that return condensed summaries of 1,000–2,000 tokens. Used in 5.1 and 5.4. Checked 2026-10-01. - Claude Docs — "Explore the context window"
Automatic compaction works like/compact,/compactwith a focus keeps what you choose,/clearbetween unrelated tasks, and delegating large reads to a subagent. Used in 5.4. Checked 2026-10-01. - Claude Docs — "How Claude Code works"
When context fills up, compaction clears older tool outputs first and then summarises the conversation. Used in 5.4. Checked 2026-10-01. - Claude Platform Docs — "Define success criteria and build evaluations"
Evals that mirror the real task distribution, edge cases and rare inputs, human-labelled test sets, and testing an LLM grader's reliability before scaling it. Used in 5.5. Checked 2026-10-01. - Claude Platform Docs — "Citations"
citations: {"enabled": true}on document blocks,cited_text, and citations "guaranteed to contain valid pointers to the provided documents". Used in 5.6. Checked 2026-10-01. - Claude Certified Architect — Foundations Exam Guide, v1.0 (Anthropic Certification Program, effective July 2026, exam code CCAR‑F).
Domain weighting, task statement definitions, and sample‑question scenarios this module is structured around.
About the practice questions
Every question in this course is an original practice item written for it, in the exam guide's own scenario style, both in the section quizzes and in the six scenario pools the mock exam draws from. None reproduces a live or retired certification item. Two mock items are close adaptations of the guide's published Sample Questions 7 and 12, and each says so in its explanation; the other sample questions are walked through in the teaching sections instead. Corrections to earlier versions of this course are listed in its errata log.
Corrections
Errata · what changed in the October 2026 revision
If you studied an earlier version, each line below gives the corrected fact for a mechanism that version taught wrong. Wording, diagram and card changes that did not change a fact are not listed.
Domain 1
- 1.1: On
stop_reason: "max_tokens", raisemax_tokensor continue the response, rather than stopping the loop. - 1.2: The coordinator can revise its decomposition through iterative refinement, so the split is not fixed once it is made.
- 1.2, 1.3:
allowedToolspre-approves tools, and the delegation tool (nowAgent) runs whether it is listed or not. The exam answer, listing"Task", is unchanged. - 1.3: Each subagent loads CLAUDE.md and its tool definitions itself, so the coordinator does not pass them on.
- 1.3: In current Claude Code, subagents run in the background by default.
- 1.4: The guide's 12% is how often the agent skips
get_customer, not how often it picks the wrong account. - 1.5: An Agent SDK callback hook that times out on
PreToolUseblocks the call. Only settings hooks of typecommand,httpormcp_toollet the call proceed, and exit code 2 is not an SDK deny route. - 1.6: Dynamic decomposition is not a synonym for the orchestrator-workers pattern, whose subtasks come from the input rather than from intermediate findings.
- 1.7:
fork_sessionis an Agent SDK option used withresume, and SDKresumetakes a session ID, while resuming by name is a CLI feature.
Domain 2
- 2.1: Anthropic's docs recommend consolidating related operations into one tool with an
actionparameter, and a misquote of its tool-writing guidance is corrected. - 2.2: Under MCP 2025-11-25, an input-validation failure is a tool execution error with
isError: true, and a client may pass protocol errors to the model. - 2.2: Error metadata such as
errorCategorybelongs inside the text content the model reads, and a validation error means "fix the argument, then retry", not "retryable". - 2.3:
tool_choice: "any"is for a required tool call when the right tool depends on the input. It and a forced tool return a 400 on the newest models and with manual extended thinking, although the exam answer is unchanged. - 2.4: An unset
${VAR}with no default is sent as the literal text${VAR}. - 2.4: When servers share a name, the whole entry from the highest-precedence scope wins, with no field merging.
- 2.4: MCP tool search is on by default, so only tool names load at session start.
- 2.5: Glob and Grep are absent from Claude Code's default tool set on macOS, Linux and WSL, although the exam answer (Grep for contents, Glob for paths) is unchanged.
Domain 3
- 3.1: A root CLAUDE.md that imports every package file still loads all of them, and per-package files save context only because other packages' files are not in the walk.
- 3.1:
/etc/claude-code/CLAUDE.mdis the managed-policy path on Linux and WSL only. - 3.2: A personal variant's different name keeps both versions for you, but it does not protect teammates.
- 3.2:
allowed-tools: Writedoes not block Bash, whiledisallowed-tools: Bashdoes. - 3.2, 3.3: Skills also accept
paths, so a glob rule is not the only way to match files by type. - 3.3: A path rule loads when Claude reads a matching file, although the guide says "editing".
- 3.4: Plan mode runs shell commands to explore, and commands outside the read-only set ask for permission, so it is not limited to read-only commands.
- 3.5: Two input/output examples do not reliably generalise to records neither example covers.
- 3.6: Exit code 0 means the run finished, so the merge gate reads blocking findings in
structured_output, not the exit code. - 3.6: The docs recommend
--barefor CI, but it is not yet the default for-p.
Domain 4
- 4.1: On current models, a conservative or confidence instruction typically raises precision and lowers recall, so real issues drop out of the list.
- 4.3, 4.4: Strict tool use does not make invalid output impossible, because a response cut off at
max_tokensor a refusal can still break the schema. - 4.3: Making a field nullable makes
nullpossible and inventions less likely, but it does not block them. - 4.5: A 6-hour submission interval meets the SLA with no slack, rather than missing it.
- 4.6: A fork of the generating session starts with its history, so it is not an independent reviewer, and a self-check line does not catch errors reliably.
Domain 5
- 5.1: Tool results are part of the conversation history and count toward the context window.
- 5.1: The start of the input is used more reliably than the middle, but it is not beyond position effects.
- 5.5: The error rate is about one support email in seven, which is one document, not one ticket type.
- 5.5: Stratified sampling draws within each segment, and small segments are covered by a fixed per-segment allocation rather than by stratification alone.
Quizzes and the mock exam
- Every section quiz and the whole mock bank were rewritten, so old question ids and answers no longer apply.
- Three old keys were wrong: hooks-q2 said a call proceeds when a hook times out, spawn-q1 treated background subagents as wrong, and the mcperrors mock item (mcperrors-x0) said a validation error is not an
isErrorresult.
All five domains of the blueprint are covered. For a final review, work through the five flashcard decks in order, retake any quiz where you missed a question, then sit the full-length mock timed.