Architect Foundations · 27% of the exam

agentic architecture and orchestration

This is the largest domain on the exam. It is about who decides what happens next, the model or your code, and what breaks when that line is drawn in the wrong place.

The scenario this domain runs on

You are building a multi-agent research system with the Claude Agent SDK. A coordinator agent takes a research topic, decides how to break it up, and delegates each part to a specialised subagent. It gathers what the subagents return and has it turned into one comprehensive report in which every claim cites its source.

The coordinator delegates to:

web searchdocument analysissynthesisreport generation
0 errorsraised in the exam guide's own failure case. Asked about AI in the creative industries, the coordinator assigned only digital art, graphic design and photography. Every subagent succeeded and the cited report read as finished, but music, writing and film were never researched, and nothing in the report said so.

Each of the seven sections below is a different place that system can go wrong. The loop never ends, or ends too early. The work is split badly. A subagent is sent off knowing nothing. A rule that had to hold every time was written as a sentence instead of as code.

Scenario switchSections 1.4 and 1.5 move to a second scenario, the guide's Customer Support Resolution Agent. The guide writes both of them in refund-and-verification vocabulary: identity checks before financial operations, and refunds above a threshold. Each is flagged again when you get there. Sections 1.6 and 1.7 stay on the research system, but the guide's own examples for them come from its code scenarios: a 14-file pull request, a legacy codebase, and a resumed codebase investigation. Each of those sections walks the guide's examples in a flagged note of its own.

Start hereThe words this domain is built on

agentic loop
The cycle your code repeats: send the conversation, read the response, run any tools it asked for, and send again.
stop_reason
A field on every API response saying why the model stopped, which your loop reads to decide whether to send another request.
coordinator
The lead agent, which splits the query, delegates the parts and combines what comes back.
subagent
A separate Claude session with its own context window. It sees nothing of the parent conversation unless you write it into the prompt.
Task tool
The tool a coordinator calls to spawn a subagent, so the exam answer lists "Task" in its allowedTools. Current SDK docs call it Agent, and Task still works as an alias.
allowedTools
An Agent SDK option that pre-approves the tools it lists, a docs reading that 1.2 sets beside the guide's.
hook
Your code, which the Agent SDK runs at a fixed point in a tool call, such as just before it executes, so it runs on every matching call.
fork_session
The Agent SDK option that, with resume, starts a new session (a saved conversation history) from a copy of an existing one, so two approaches can grow from the same baseline.
Agent SDK
Anthropic's library for building agents on the same loop, tools and hooks as Claude Code, its command-line coding agent.
MCP
The Model Context Protocol, a standard way to connect an agent to outside tools. The Agent SDK names an MCP tool mcp__<server>__<tool>. Domain 2 covers it.
CLAUDE.md
A file of shared project instructions that Claude Code loads automatically, and that each subagent loads for itself. Domain 3 covers it.
few-shot
Worked input-and-output pairs placed in a prompt to show the model the pattern to follow. Domain 4 covers them.

task statements

videos

flashcards

0 / 21 known
Card 1 of 21