OpenClaw Multi-Agent Content Factory: Research, Writing & Thumbnail Agents

Introduction

Creators use Multi-Agent Content Factories within Discord — chaining specialized agents: a Research Agent scans for trending stories, a Writing Agent drafts scripts, and a Thumbnail Agent generates cover art. These agents communicate autonomously via sessions_spawn and sessions_send, delivering a complete content package for human review every morning at 8:00 AM.

The Content Factory pattern represents OpenClaw's strength: autonomous, coordinated workflows that produce tangible outputs. You're not asking one agent to do everything—you're building a team of specialists that hand off to each other. The result is a content pipeline that runs while you sleep. Wake up, review the package, publish or iterate. The factory runs again tomorrow.

This guide walks through the architecture, each agent's role, and how to implement your own. Whether you're a YouTuber, newsletter writer, or social media manager, the pattern adapts to your format.

Architecture

The Content Factory pattern uses three specialized OpenClaw agents:

1. Research Agent: Monitors RSS, Twitter/X, GitHub releases; identifies trending topics; outputs topic briefs with sources and suggested angles. Runs on a schedule (e.g., every 6 hours) or triggered by "enough new content."
2. Writing Agent: Receives briefs; drafts scripts/articles; applies tone and format guidelines from SOUL.md. Output is structured (word count, subheadings, CTA) and ready for the next stage.
3. Thumbnail Agent: Receives script summaries; generates cover art via image API (DALL-E, Midjourney, or local Stable Diffusion); outputs final assets with consistent naming. Posts to Discord for human review.

Agents share memory via Markdown files. Research writes to TOPICS.md; Writing reads TOPICS.md and writes DRAFTS.md; Thumbnail reads DRAFTS.md and produces IMAGES/. The file system is the handoff mechanism. No complex message queues—just files that the next agent reads.

The beauty of this design: each agent can be tuned independently. Research gets different HEARTBEAT.md than Writing. Each has its own SOUL.md for domain-specific behavior. The Research Agent is concise and source-focused; the Writing Agent is creative and brand-aware; the Thumbnail Agent is visual and format-specific. Specialization improves quality.

Research Agent

Configured with HEARTBEAT.md: "Every 6 hours, scan [list of 100+ sources]. Deduplicate by title similarity. Apply quality score (Priority Source +3, Recency +2). Write top 5 to TOPICS.md."

Uses web search skill, RSS parsing, and deduplication logic. Output: structured topic list with sources, relevance scores, and suggested angles. The Research Agent doesn't write—it curates. It answers "what's worth covering?" so the Writing Agent can focus on "how do we cover it?"

Source configuration matters. Tech creators might prioritize Hacker News, TechCrunch, and GitHub trending. Finance creators might prioritize earnings calendars, SEC filings, and financial news. The Research Agent's SOUL.md encodes domain expertise: "For tech, prioritize shipping and launches. For finance, prioritize guidance and surprises."

Deduplication is critical. The same story appears on 10 sites. The Research Agent clusters by semantic similarity and picks the best source. Without it, you'd get 5 variations of the same news. With it, you get 5 distinct stories.

Writing Agent

Triggered by Research completion (file watcher or heartbeat dependency). Reads TOPICS.md, selects highest-priority topic, drafts full script. Uses SOUL.md for brand voice and format (e.g., "800 words, 3 subheadings, conclusion with CTA"). Writes to DRAFTS.md and notifies Thumbnail Agent.

The Writing Agent is where creativity lives. It takes a dry topic brief and turns it into engaging content. The SOUL.md might specify: "Conversational tone. Use analogies. End with a question to drive engagement. Avoid jargon—explain technical terms." The agent applies these consistently across drafts.

For video scripts, the Writing Agent might include timestamps, hook suggestions, and B-roll notes. For articles, it might include meta descriptions and suggested images. The output format depends on your content type—the agent adapts.

Thumbnail Agent

Receives draft summary. Calls image generation API (DALL-E, Midjourney API, or local Stable Diffusion) with prompt derived from draft. Saves to IMAGES/ with consistent naming. Posts to Discord channel for human review.

The Thumbnail Agent's job is to create visuals that match the content. It reads the draft summary, extracts key themes and emotions, and generates a prompt. "Tech product launch, excited crowd, modern aesthetic, bold text overlay" might become a DALL-E prompt. The agent can generate 2-3 options for A/B testing.

Naming convention matters. 2026-02-19_topic-slug_v1.png keeps things organized. The agent writes a brief to DRAFTS.md: "Thumbnail generated. Key elements: X, Y, Z. Suggested headline: ..." The human has context when reviewing.

Orchestration

sessions_spawn and sessions_send enable inter-agent communication. Research Agent completes → spawns Writing Agent with TOPICS.md path. Writing completes → spawns Thumbnail Agent with DRAFTS.md path. Final output lands in Discord by 8 AM.

Human reviews the package: topic selection, script quality, thumbnail appeal. Approves or requests revisions. The factory runs again the next cycle. Revisions can be handled by re-triggering the relevant agent with feedback—"Writing Agent: revise draft for more technical depth."

The orchestration can be time-based (Research at 2 AM, Writing at 4 AM, Thumbnail at 6 AM) or event-based (Research completes → spawn Writing). Time-based is simpler; event-based is faster. For daily content, time-based usually suffices.

Implementation

• Run all three agents on same host; shared ~/clawd/ memory directory. They need to read each other's output files.
• Use HEARTBEAT.md with time conditions: Research at 2 AM, Writing at 4 AM, Thumbnail at 6 AM. Adjust for your timezone and schedule.
• Discord webhook for final delivery. The Thumbnail Agent (or a final "Publisher" agent) posts the package to a channel. You get a notification; you review on your phone.
• Cost: ~$20-40/month in API fees for daily content cycle. Research is cheap (mostly retrieval); Writing and Thumbnail consume more tokens. Image generation adds cost. Use two-tier processing where possible—deterministic steps before LLM.

Start with a weekly cycle before going daily. Validate each agent's output quality. Tune SOUL.md and HEARTBEAT.md. Then increase frequency. Many creators run 2-3x per week—enough for a steady stream without overwhelming the review process.

Costs and Optimization

API costs scale with volume. A daily cycle might use: Research ~50K tokens, Writing ~100K tokens, Thumbnail ~20K tokens + image gen. At $3/M tokens (Claude) and $0.04/image (DALL-E), that's roughly $0.50-1.00 per day—$15-30/month. Add image variations or more drafts, and costs rise.

Optimization strategies: Use smaller/cheaper models for Research (it's mostly retrieval and summarization). Reserve premium models for Writing (quality matters). Batch image generation—generate multiple thumbnails in one call. Use two-tier processing: Research's deduplication can be script-based; only the "pick top 5" needs the LLM.

Wrapping Up

The Multi-Agent Content Factory demonstrates OpenClaw's strength: autonomous, coordinated workflows that produce tangible outputs. It's not a demo—it's a production pattern used by creators today. Start with one agent (Research), add Writing, add Thumbnail. Iterate on quality. Scale the schedule. See multi-agent patterns and Discord setup for implementation guides.