Capstone 7: A Multi-Agent Content Pipeline

Ask one model call to "research and write a great article" and you get something fluent that nobody checked. Professional publishing works differently: a researcher gathers facts, a writer drafts from them, a fact-checker verifies every claim, an editor sends it back until it's good, and a human signs off before anything goes out. Each role catches what the others miss.

In this capstone you'll build that newsroom as a pipeline of specialized agents. They pass structured handoffs instead of chat transcripts, code verifies every fact's quote against its source, an evaluator–optimizer loop revises the draft until it meets a bar defined in code, and nothing publishes until a person approves the exact text. It runs on a mock source pack, so you only need an Anthropic API key.

What you'll build#

A command-line pipeline that turns a brief into a publish-ready article:

npx tsx main.ts

It prints each stage as it runs, shows you the final draft with its check results and any unresolved issues, and asks you to approve, reject, or request changes. Every handoff and every revision round is saved to out/ as JSON, so you can audit what happened or rerun a later stage without paying for earlier ones.

The example brief is a customer announcement for offline mode in Acme Planner, the invented product from Capstone 2. The sources include an internal roadmap note that must never reach customers, which gives your guardrails something real to catch.

This capstone combines Handoffs and Shared State, Evaluator–Optimizer Loops, the grounding approach from Capstone 1, the editorial workflow from Writing and Content Workflows, and approval design from Human-in-the-Loop Design.