Building an Autonomous Content Pipeline with Claude Code

The Pipeline

I built an autonomous content pipeline using Claude Code. The system runs three phases in sequence: Observer, Strategist, and Marketer.

Each phase is an independent Claude session that reads the previous phase’s output and generates the next step.

What Broke

After 6 rounds of testing, I found 9 bugs:

• Parallel execution conflicts: Cron jobs fired all 3 phases simultaneously. The Strategist hadn’t finished when the Marketer started.
• Theme duplication: Without an exclusion list, the pipeline kept picking the same topic.
• Self-reported quality checks: The AI was checking its own work and always passing itself.

Every single bug was in the harness — the environment around the model — not in the model itself.

The Fix

I switched from time-based cron to event-driven chaining with after dependencies. Each phase only starts when the previous one completes.

# Before: all fired at the same time
observer: "0 7 * * 1"
strategist: "0 7 * * 1"
marketer: "0 7 * * 1"

# After: event-driven chain
observer: "0 7 * * 1"
strategist:
  after: observer
marketer:
  after: strategist

Takeaway

AI agent quality is determined outside the AI. The model is the chef — but if the kitchen is broken, no chef can cook.

---

Want to go deeper?

This article shows one piece. The full system — five vendor interpretations, six building blocks, and the implementation patterns from AGENTS.md to Self-Evolving Agents — is in Harness Engineering — From Using AI to Controlling AI.