Capstone: A Production-Style Multi-Agent Pipeline for website-content

Fundamentals' own capstone mapped seven agent types to seven real steps in generating a course chapter for website-content, run one at a time, by hand. Its own tip box named this exact capstone as where that gets formalized into a real, chained pipeline. This capstone builds it — applying every chapter of this course to turn that manual workflow into an automated one.

The Scenario

website-content generates course chapters repeatedly and often — this pipeline exists specifically because that repeated volume (Chapter 9's own deciding question) genuinely justifies building it, unlike a one-off task. The goal: take a chapter-generation request in and produce a rules-compliant, cross-reference-accurate chapter out, with tracking files updated automatically, and a human reviewing only what genuinely needs judgment.

Stage 1 — Fan-Out Research (Chapter 3)

Three research agents launch in parallel: one checking course_folder_mapping.md for the correct folder and filename convention, one checking course_name_index.md for the full course name, and one checking completed_courses.md for any existing entry that might overlap. These three checks are genuinely independent of each other — a real fan-out, per Chapter 3's own deciding question.

Stage 2 — Fan-In Planning (Chapter 6)

Once all three research agents finish, a planning agent proposes the chapter's actual outline, informed by their combined findings — this step genuinely cannot start until every one of the three parallel results is available, the fan-in half of Chapter 3's pattern. The plan itself stays a reviewable proposal, not an implementation, per Chapter 6.

Carrying Constraints Through Shared State (Chapter 7)

The original request's own hard constraints — the required P2 banner format, the P16 code-block/copy-button convention, this course's established accent color and wrapper class — are added to shared state at the very start, not just handed forward step to step. This guarantees they're still visible at the final review stage, several hops downstream, rather than risking silently dropping out along the way.

Stage 3 — Generation (Fundamentals Chapter 5)

A coding agent, briefed with the approved plan and the shared-state constraints together, writes the actual chapter — matching this project's own established per-course template exactly, the same way every chapter in both courses of this track was actually produced.

Stage 4 — Rules-Compliance Review (Fundamentals Chapter 6, specialized per Chapter 5)

A review agent — read-only, no Edit — checks the generated chapter against the shared-state constraints directly: is the banner present and correctly formatted, does every code block use the copy-button markup, is the wrapper class and accent consistent. This is the same specialized-review-agent pattern this course's own Chapter 5 applied to security — here applied to rules compliance instead.

Handling a Failure at Review (Chapter 8)

If the review stage finds a violation — a missing banner field, say — debugging means checking whether the plan itself specified the right convention in the first place, or whether the generation stage simply failed to follow a plan that was actually correct, per Chapter 8's own "check the earliest step first" guidance, rather than assuming the generation agent is automatically at fault.

Was This Pipeline Worth Building? (Chapter 9)

Applying Chapter 9's own deciding question honestly: website-content generates chapters at real, ongoing volume — not once or twice, but repeatedly across many sessions. This is exactly the profile Chapter 9 names as justifying a pipeline's fixed setup cost, rather than a case where a direct approach would have been more efficient.

Running Autonomously (Chapter 4)

Once built, this pipeline doesn't need to be launched by hand each time — it can run as a background process, triggered whenever a new bucket-list entry is marked ready, checking back on its own progress at a cadence matched to how long generation and review actually take, per Chapter 4's own wake-up guidance, rather than polling constantly or requiring manual restarts.

Hands-On Exercises