How to Build a Multi-Agent SEO Pipeline (Case Study)

We built a fully automated 7-agent content pipeline — from keyword research to published article — running on a single $10 VPS. Here's the exact architecture, agent breakdown, costs, and what we learned. This is 100% dogfood: we're using it to build this very site.

Table of Contents

• The Problem
• Architecture Overview
• 1. SERP Agent — Keyword Discovery
• 2. Source Agent — Research Collection
• 3. Brief Agent — Content Planning
• 4. Draft Agent — Article Writing
• 5. Visual Agent — Diagrams & Images
• 6. Doc Agent — Formatting & Publication
• 7. Distribution Agent — Promotion
• Orchestration with Hermes Agent
• Cost Breakdown
• Early Results
• How to Replicate This

The Problem

Content marketing at scale is broken. A single high-quality technical article takes 6-12 hours of research, writing, editing, formatting, and promotion. Most teams either publish infrequently or sacrifice quality for volume.

We wanted a different approach: automate the pipeline, not the quality. Use AI agents to handle research, drafting, formatting, and promotion — while humans stay in the loop for editorial decisions, strategy, and quality control.

This case study documents exactly how we built a 7-agent SEO content pipeline using Hermes Agent as the orchestration layer. The pipeline runs on a single VPS, costs under $100/month to operate, and can produce a research-backed, SEO-optimized article from concept to publication in under 2 hours of wall clock time.

Architecture Overview

Each agent is a Hermes Agent skill — a reusable, parameterized workflow that can be invoked independently or chained as part of a larger pipeline. The orchestrator (built with Hermes' Kanban and cron systems) manages dependencies, parallel execution, and retry logic.

Tech Stack

1. SERP Agent — Keyword Discovery

The pipeline starts with research. The SERP Agent performs automated keyword discovery across multiple dimensions:

• Search autocomplete — extracts "people also search for" suggestions from search engines
• Related searches — scrapes "related searches" sections to build keyword clusters
• Trend analysis — identifies rising queries in the niche using trend data
• Competitor gap analysis — identifies keywords competitors rank for that we don't

Output: A structured keyword brief with search intent, estimated competition, and recommended article topics. This feeds directly into the content calendar.

2. Source Agent — Research Collection

Once a topic is selected, the Source Agent gathers authoritative references:

• Official documentation and GitHub repos
• Recent blog posts and technical articles
• Community discussions (Reddit, Discord, Hacker News)
• Benchmark data and comparison tables
• Statistic and market data from industry reports

Output: A curated source document with URLs, key quotes, statistics, and credibility scores. The Brief Agent uses this to build the article outline.

3. Brief Agent — Content Planning

The Brief Agent synthesizes keyword research and source data into a structured content brief:

• Target keyword placement map (H1, H2, meta, first 100 words)
• Article outline with section-level keyword targeting
• Source citations mapped to sections
• Internal linking opportunities
• CTA and affiliate placement suggestions

Gate: This is the human review step. The brief is reviewed and approved before the Draft Agent writes the article. Quality control happens here, not after 2,000 words have been written.

4. Draft Agent — Article Writing

The Draft Agent generates the full article from the approved brief:

• Generates section-by-section following the outline
• Incorporates source citations and quotes inline
• Optimizes for target keywords naturally (no keyword stuffing)
• Generates meta description, alt text suggestions, and excerpt
• Writes in the site's established tone (technical, honest, authoritative)

Quality controls: Every claim is tagged with its source. The Draft Agent has access to the Source Agent's research document and must cite facts. If a source is unavailable for a claim, it flags it as "needs verification" rather than inventing a citation.

5. Visual Agent — Diagrams & Images

The Visual Agent creates supporting graphics for each article:

• Architecture diagrams (SVG)
• Comparison tables (HTML with CSS styling)
• Flowcharts for process-oriented articles
• Social media cards (1200×630px Open Graph images)

For our dark-themed tech site, visuals use a consistent color palette (green accent on dark backgrounds, matching the site theme).

6. Doc Agent — Formatting & Publication

The Doc Agent handles everything between "article written" and "article live":

• Converts markdown to production HTML with proper semantic structure
• Injects Schema.org structured data (Article, FAQ, HowTo as appropriate)
• Optimizes images and generates responsive srcset
• Updates the sitemap.xml
• Creates the article page under /articles/{slug}/
• Updates the homepage "Latest Articles" section
• Pings search engines for indexing

Output: A fully published, SEO-optimized article page on qantcore.space.

7. Distribution Agent — Promotion

The final agent pushes the article into distribution channels:

• Social media posts (X/Twitter, LinkedIn, relevant subreddits)
• RSS feed update and ping
• Email notification to subscribers
• Discord community post
• Scheduled re-promotion (t + 7 days, t + 30 days)

Analytics feedback: The Distribution Agent also monitors initial engagement (clicks, impressions, shares) and reports back to the SERP Agent for iterative keyword refinement.

Orchestration with Hermes Agent

The magic is in the orchestration. Each agent is a Hermes skill — a SKILL.md file with YAML frontmatter describing the agent's behavior, inputs, outputs, and tool requirements. The orchestrator uses Hermes' native capabilities:

Kanban Orchestrator

Hermes has a built-in Kanban-based task decomposition engine. For each article, the orchestrator:

• Creates a workflow with 7 steps (one per agent)
• Manages dependencies (e.g., Draft cannot start until Brief is approved)
• Allows parallel execution where possible (SERP can run while Source is collecting for another article)
• Provides visibility into pipeline status via the TUI

Cron Scheduling

Recurring tasks are scheduled via Hermes' cronjob system:

• Weekly: SERP Agent runs to discover new keyword opportunities
• Per-article: Source → Brief → Draft → Visual → Doc → Distribution, each triggered by the previous completion
• Bi-weekly: Distribution Agent re-promotes older content
• Monthly: Full content audit and refresh cycle

Parallel Subagent Execution

With delegate_task, agents can run in parallel. For example, after the Brief is approved, the Draft Agent and Visual Agent can work simultaneously — the diagram doesn't need the full article text, just the outline.

Cost Breakdown

Compare this to the cost of a freelance technical writer ($200-500/article) or the time cost of writing in-house (6-12 hours × $50-100/hr = $300-1200). Even with human review time (30 minutes per brief), the pipeline delivers a 10-50× cost reduction.

Early Results

As of May 2026, the pipeline is live and producing content for qantcore.space. Early metrics:

• Setup time: 2 days from idea to first published article
• Articles produced: 1 published, 2 in pipeline
• Time per article: ~2 hours wall clock (30 min human review, 90 min automated)
• Cost per article: ~$1.86 in infrastructure + API costs

We'll update this section as more data comes in. Check back for traffic, rankings, and conversion metrics.

How to Replicate This

Want to build your own multi-agent SEO pipeline? Here's the high-level roadmap:

1. Install Hermes Agent — Follow our Hermes Agent Setup Guide (coming soon) for a zero-to-running installation on any Linux VPS.
2. Configure tools — Set up web search provider, file system access, and LLM provider in config.yaml.
3. Create skills — Write SKILL.md files for each agent in your pipeline. Start with SERP and Source, then build up.
4. Test the loop — Run a single article end-to-end manually. Identify bottlenecks and refine prompts.
5. Schedule — Set up cron jobs for recurring execution. Start with weekly SERP runs and build from there.
6. Iterate — Monitor output quality, adjust agent prompts, and add new agents gradually.

The full pipeline configuration (skills, prompts, cron schedules) will be open-sourced once stable. Follow QantCore for updates.

Pro Tips

• Start small: A 2-agent pipeline (SERP + Draft) is better than a broken 7-agent one. Add complexity as you validate each stage.
• Human gate matters: The Brief review step is not optional. It's where quality is controlled before the expensive (in token terms) drafting step.
• Monitor costs: Each agent call costs tokens. The SERP + Source + Brief + Draft pipeline for this article cost about $1.50 in API calls. Track it.
• Use skills: Hermes' skill system is the backbone of reusability. A well-tuned skill can be shared across articles and even across different sites.