The Concept: Multiple AI Calls, Specific Roles
Session 9.1 · ~5 min read
One Call vs. Many
Up to this point, your pipeline uses a single AI call for draft generation. One prompt, one system message, one response. That works for simple content. For complex content, a single call tries to do too much: research, organize, write, maintain voice, check facts, and format, all in one generation. The result is mediocre at everything instead of excellent at any one thing.
A multi-agent workflow splits this workload across multiple AI calls, each with a specific, narrow role. Agent 1 researches. Agent 2 outlines (if you choose to automate this). Agent 3 writes. Agent 4 edits. Agent 5 formats. Each agent has its own system prompt, optimized for its specific task.
This is not a conversation. It is a production line. Each agent receives structured input, produces structured output, and passes it forward. No context is shared implicitly. Everything is explicit and documented.
Why Specialization Improves Output
Consider a system prompt that tries to do everything: "You are a researcher and writer who produces well-sourced, voice-consistent articles in my style with no AI artifacts, formatted in markdown with proper headings." That prompt pulls the model in five directions simultaneously. Voice preservation competes with factual accuracy for attention in the context window. Formatting competes with argumentation.
Now consider three separate agents:
'Find and verify sources'"] --> B["Writing Agent
'Draft in this voice'"] B --> C["Editing Agent
'Flag issues per rubric'"] style A fill:#222221,stroke:#c8a882,color:#ede9e3 style B fill:#222221,stroke:#6b8f71,color:#ede9e3 style C fill:#222221,stroke:#c47a5a,color:#ede9e3
Each agent's system prompt is focused. The Research Agent does not care about voice. The Writing Agent does not care about source verification. The Editing Agent does not care about prose generation. Each does one thing well.
| Approach | Context Window Usage | System Prompt Focus | Output Quality |
|---|---|---|---|
| Single agent (everything in one call) | Overloaded with competing instructions | Diluted across multiple objectives | Mediocre at everything |
| Multi-agent (specialized calls) | Each call uses context efficiently | Laser-focused on one task | Strong at each stage |
The Minimum Viable Agent Chain
The simplest useful agent chain has three agents: Research, Writer, Editor. This maps directly to the pipeline stages from Module 8 but automates the transitions between them.
Agent 1: Research Agent. System prompt focuses on information gathering, source evaluation, and structured output. Input: a topic and research questions. Output: a structured research brief (JSON or Markdown with defined sections).
Agent 2: Writing Agent. System prompt contains your voice fingerprint and writing constraints. Input: the research brief from Agent 1, plus your outline and content spec. Output: a first draft.
Agent 3: Editing Agent. System prompt contains your quality rubric and the 15 forensic markers. Input: the draft from Agent 2. Output: an annotated version with issues flagged and scored.
Human review happens after Agent 3, not before. The Editing Agent does not replace human review. It pre-screens the draft so your human review time is spent on genuine judgment calls rather than catching obvious issues.
Agent System Prompts
Each agent needs a system prompt engineered for its specific role. Generic system prompts defeat the purpose.
| Agent | System Prompt Focus | Key Instructions |
|---|---|---|
| Research | Information accuracy and structure | Only report verified facts. Cite every claim. Flag uncertainty. Output as structured JSON. |
| Writer | Voice preservation and argument | Follow the outline exactly. Use only provided research. Match voice fingerprint. No hedging. |
| Editor | Quality detection | Score each dimension 0-10. Flag every AI artifact. Mark unverified claims. Do not rewrite. |
A multi-agent workflow is not about having more AI. It is about having more focused AI. Each agent does less, but does it better. The quality improvement comes from specialization, not from additional computation.
When Single Agent Is Still Fine
Multi-agent workflows add complexity. That complexity is justified for production content that must meet high standards at scale. It is not justified for every task.
- Single agent: short content (under 500 words), brainstorming, quick internal communications, social media posts
- Multi-agent: long-form articles, course content, book chapters, product descriptions at scale, anything that will be published under your name and persist on the internet
The deciding factor is accountability. If nobody will trace the content back to you, a single agent is fine. If your name is on it, the chain is worth the extra cost and complexity.
Further Reading
- Multi-Agent Workflows: A Practical Guide, Kanerika (Medium)
- Multi-Agent Content Creation System: Complete Guide, Sight AI
- Building Autonomous Systems: A Guide to Agentic AI, DigitalOcean
Assignment
Design a 3-agent chain for a content type you produce regularly:
- Define Agent 1 (Researcher): system prompt, expected input format, expected output format
- Define Agent 2 (Writer): system prompt (include your voice fingerprint), expected input, expected output
- Define Agent 3 (Editor): system prompt (include your quality rubric), expected input, expected output
Write each system prompt in full. Do not summarize. The system prompt is the agent's operating instructions. Test each agent individually by running it in isolation with appropriate input. Does each agent produce output that matches its role?