Agentic Architecture & Orchestration
One Claude call isn't enough. Build a multi-agent reviewer on the Claude Agent SDK.
9 lessons 3 phases advanced
Get started
Signed in as your account.
why this workshop
who it's for.
Developers who finished W3 and are ready to build a multi-agent reviewer on the Claude Agent SDK.
exam coverage
Maps to Domain 1 (Agentic Architecture & Orchestration), 27% of the CCA – Foundations exam — its largest domain.
what you'll build
what you walk away with.
- 01 A `runAgentQuery()` SDK-loop runner beside the hand-rolled loop, plus `models.ts` pinning full model IDs (Sonnet 4.6 coordinator, Haiku 4.5 workers and critic).
- 02 A sequential coordinator that decomposes the diff locally and spawns a file-reviewer subagent per file via the Agent tool — its AgentDefinition inline, for now.
- 03 The reviewer extracted into `worker.ts`, and the coordinator flipped to parallel Agent fan-out — the SDK runs the subagents concurrently; your prompt caps concurrency at 5.
- 04 An aggregator: dedupe and severity ranking in pure code, a findings-critic model pass, and a resume-based coverage-gap refinement round.
- 05 PreToolUse and PostToolUse hooks — a cost-cap deny that steers the model with its reason, a secrets policy gate, and a JSONL audit log of every tool call.
- 06 Session machinery: capture the session id off the init message, resume the coordinator with full prior context, fork divergent branches from a shared baseline.
- 07 Four decomposition strategies behind a `--strategy` flag: fan-out-file, fan-out-concern, pipeline, single.
- 08 SDK-native failure handling: branch on result subtypes, salvage findings from completed workers, report structured errors[] — no hand-rolled retry/backoff.
- 09 Schema-enforced output (`outputFormat: json_schema`) plus a per-model cost report on the coordinator — then the full system run against the 30-file fixture.
curriculum
9 lessons across 3 phases.
phase A · Agent foundations 3 lessons
- 01 Agent SDK Intro
- 02 Orchestrator: Decompose by File
- 03 Worker Agents
phase B · Production patterns 3 lessons
- 04 Aggregation + Critic
- 05 Hooks
- 06 Sessions
phase C · Hardening 3 lessons
- 07 Task Decomposition Patterns
- 08 Failure Modes
- 09 End-to-End on 30 Files
prerequisites
what you need before you start.
- Completed W3 (or equivalent)
- The workshop reads src/review.ts as its starting state — reviewToolDriven() must be present. The MCP server (mcp/server.ts) must exist and be registered in .mcp.json.
- Node.js 22+
- Same stack as W1–W3 — pnpm workspaces, tsx, TypeScript. `pnpm install` brings the Agent SDK and its bundled runtime via the repo's lockfile — no separate Claude Code install.
- git
- Used by fixture generation scripts. Standard install.
start
how to start.
Once Claude Code is set up, tell it “start the agentic workshop” and it clones the project and queues the first lesson.
New here? The
getting-started guide
walks you through installing the lwc CLI and the Claude Code
plugin once — everything you need before that line works.
the series