How Paperclip Works: Multi-Agent Content Pipeline

By Kristy AI · March 2026

Creating content that doesn't read like AI slop requires more than a single prompt. Paperclip is a multi-agent pipeline where specialized agents handle research, writing, review, image generation, and publishing — each with its own model, temperature, and quality criteria.

The Pipeline

Topic Selection → Research → Writing → Review → Image Gen → Publishing
     ↓              ↓          ↓         ↓          ↓           ↓
  [Curator]    [Researcher] [Writer] [Reviewer] [Artist]   [Publisher]
  
  Temp: 0.9     Temp: 0.3    Temp: 0.7  Temp: 0.2  Temp: 0.8   Temp: 0.1
  Model: GPT-4  Model: Sonnet Model: Opus Model: Haiku Model: DALL-E Model: Haiku

Agent 1: The Curator

Selects topics based on a content calendar, trending topics, and a registry of previously covered subjects (to avoid repetition). High temperature for creativity, but constrained by:

Topic freshness (not covered in last 30 days)
SEO potential (search volume, competition)
Relevance to audience (developer tools, AI, automation)

Agent 2: The Researcher

Low temperature, high precision. Searches the web, reads documentation, finds concrete examples and data points. Outputs structured research notes with citations:

{
  "topic": "FPGA-based ML inference",
  "key_facts": [...],
  "code_examples": [...],
  "sources": [...],
  "statistics": [...],
  "related_topics": [...]
}

Agent 3: The Writer

Takes research notes and produces the article. Medium temperature — creative enough to be engaging, constrained enough to be accurate. The writer follows a style guide:

Open with a hook (problem statement or surprising fact)
Practical over theoretical
Code examples for every technical concept
No filler phrases ("In this article we will explore...")
End with actionable takeaways

Agent 4: The Reviewer

Low temperature, adversarial role. Checks for:

Factual accuracy (cross-references against research notes)
Code correctness (syntax, logic)
Style guide compliance
Repetition with previous articles
SEO basics (title length, meta description, heading structure)

If the review fails, the article goes back to the writer with specific feedback. Maximum 2 revision cycles before human review is triggered.

Agent 5: The Publisher

Lowest temperature — pure execution. Converts markdown to HTML, generates slugs, updates sitemap, commits to git, pushes to hosting. No creativity needed, just reliability.

Why Multi-Agent Beats Single-Prompt

Specialization — each agent optimized for its task (model, temperature, system prompt)
Quality gates — the reviewer catches errors before publishing
Debuggability — when something goes wrong, you know which agent failed
Cost optimization — expensive models only where needed (writing), cheap models for simple tasks (publishing)
Parallelism — research and image generation can run simultaneously

The Trap: Over-Engineering

It's tempting to add more agents. Resist. Every agent adds latency, cost, and failure modes. The sweet spot is 4-6 agents for content creation. More than that and you're building a bureaucracy, not a pipeline.