joshualarosa.ai
ResourcesPodcastNewsletterWork with me
← ResourcesMarketing

20 AI Agents That Write Viral Launch Videos

How Shown Media uses a multi-agent system to turn a single brand brief into a shoot-ready script


One Brand Brief In, Shoot-Ready Script Out

Shown Media built a 20-agent AI system that researches, writes, reviews, and stress-tests every launch video script they produce. Their launches average 2 million views per video. One client alone drove over $3 million in revenue. Across their portfolio, they have 50 million+ total views.

Most people open ChatGPT, type a prompt, and take whatever comes back. Shown Media called that garbage. So they built something different.

Each of the 20 agents has exactly one job, a strict checklist, and a default position of REJECT. The system runs in five phases, from raw keyword research through final delivery. The output is a complete, production-ready script with multiple hook options, a full demo body, and CTA variations.

Key Takeaway

Every script goes through at least 6 adversarial review gates, checked against real platform data from YouTube, X, and Reddit, stress-tested for clarity, and verified against the founder's own words from the intake call. The system's default is to reject, not approve. What survives is ready to perform.

This guide breaks down every phase, every agent, and exactly what each one does.

The Research Agent Builds the Foundation

Agent 1 / Research Agent

Everything starts here. The Research Agent takes the brand brief and generates 50 keywords across 7 categories: pain points, product features, competitors, industry trends, audience language, use cases, and outcomes.

Each keyword gets scored on viral potential and novelty. The top 15 become the seed file that every downstream research agent pulls from.

The keyword scoring is what separates this from a basic brainstorm. Instead of guessing which angles might work, the system ranks them by measurable potential before a single search happens.

What it outputs: A ranked keyword list (top 15) that drives all research in Phase 2.

Four Agents Mining Four Platforms in Parallel

All four research agents run at the same time. Each one mines a different platform for different types of ammunition. They are not looking for content to reference. They are reverse-engineering what language actually performs right now.

Agent 2 / YouTube Research Agent

Runs three Playwright passes on YouTube using the top 15 keywords: default sorting, filtered to the last 12 months, and filtered to the last month. For each search, it plays a "Ceiling/Floor Game" — comparing high-view titles against low-view titles to extract what copy patterns are earning clicks.

Agent 3 / X (Twitter) Research Agent

Pulls roughly 5,000 tweets across five searches: viral brand mentions, competitor frustration, trending conversations, customer pain language, and launch post formats. Every tweet gets tagged by function: WEAPON, HOOK-AMMO, BODY-AMMO, CTA-AMMO, or CONTROVERSY-AMMO.

Agent 4 / Reddit Research Agent

Mines the top 6 subreddit threads for high-upvote comments about frustrations, failed alternatives, and "I'm about to quit" scenarios. Every quote gets an emotion tag (FRUSTRATED, DESPERATE, ANGRY, HOPEFUL) and a usage note explaining how it can be deployed in the script.

Agent 5 / Industry Research Agent

Hits competitor pricing pages, G2 and Trustpilot review pages, and industry stat sources. Extracts exact pricing tiers, hidden fees, top complaints, and market-level numbers with full source attribution.

Agent 6 / Research Compiler Agent

Waits for all four research agents to finish, then synthesizes everything into a 50-to-100-nugget base organized by what each downstream writer needs: hook ammo, body ammo, proof ammo, CTA ammo, and controversy ammo. Nothing gets invented. The compiler sorts and prioritizes what the research already found.

Key Takeaway

The research phase produces real language from real people on real platforms. No AI-generated assumptions. Every nugget traces back to a tweet, a Reddit comment, a YouTube title, or a verified stat.

Writing and War-Testing Every Hook

The hook pipeline and the giveaway pipeline run in parallel. Each goes through its own adversarial review cycle.

Agent 7 / Hook Writer

Writes 5 distinct hooks, each built on a different framework: Standard Raise, Visual Demo, Mission-Based, Creative Format, and Wild Card. Each framework gets filled with brand-specific details pulled from the brief and the nugget base.

Agent 8 / Hook Manager

The adversarial gatekeeper. Reviews all 5 hooks against a 10-point binary checklist:

  • Does it contain a specific number?
  • Does it use a violent verb?
  • Does it name an enemy?
  • Is it under the character limit?
  • Does it avoid every banned word?
  • Is it specific to this brand, not a generic template?

Then it runs a Library Dominance Test, comparing each hook against their best-performing historical hooks. Default position is REJECT. The writer gets up to 3 rounds of revisions with surgical instructions.

Agent 9 / Giveaway Writer

Writes 2 CTA options following the Sacred CTA Flow: named giveaway, specific contents, proof stat, dollar outcome, platform-specific trigger words.

Agent 10 / Giveaway Manager

Pulls real high-performing giveaway posts from X as live benchmarks, then adversarially reviews both CTA options against a 15-point checklist. Also runs a Library Dominance Test and an X Reference Comparison Test. Default: REJECT. Max 3 revision rounds.

The Hook Agent alone writes 4 hooks with 3 full rewrites each. Nothing moves forward until it scores 10/10. This is why their hooks consistently outperform — they are tested against real data before anyone ever films.

Seven Agents Refine Every Line of the Script

This phase runs sequentially. Each agent receives the body from the previous one and makes targeted upgrades before passing it forward.

Agent 11 / Body Writer Writes the full demo walkthrough: 7+ narrated steps, 2+ "intelligence moments," and a specific business scenario grounding the demo. Agent 12 / Weapons Specialist Upgrades 2-4 lines using Economy Killers, Savage Comparisons, Subtle Controversy, Mission Alignment, Technical Flex, and Confidence Bets. Agent 13 / Controversy Specialist Inserts 1-2 subtle controversy lines tied to who loses when the product wins. Tests for screenshot-worthiness and taste. Agent 14 / Technical Specialist Upgrades every technical line to name specific technologies, data sources, and counts. No vague language allowed. Agent 15 / Flow Specialist Scans for AI-written tells: repeated starters, fake drama, banned transitions, corporate filler, passive voice, and hedging. Agent 16 / Body Manager Final adversarial gate: 15-point checklist, Library Dominance Test against the 2 best library bodies. Default: REJECT.

Agent 17 / Fathom Checker (Brand Advocate)

Re-reads the full client transcript from the intake call and acts as the brand's internal advocate. Default position is REJECT. It asks: Does this script capture what actually makes the product special? Does it use the founder's exact phrases? Does it reflect what the client emphasized on the call?

The Fathom Checker does not write. It only routes surgical fix instructions back to the specific agents who need to make changes. This keeps the brand voice authentic to the founder, not the AI.

Three Final Gates Before Delivery

Agents 18 and 19 run in parallel. Agent 20 runs last.

Agent 18 / Mom Test

A binary clarity check on every sentence. Flags anything a non-technical 55-year-old would not understand: jargon, acronyms, industry shorthand, assumed knowledge. Rewrites flagged sentences using plain-language patterns.

Agent 19 / Call Supervisor

Checks the full assembled script against the Fathom transcript one more time: feature coverage, emphasis alignment, hook-to-body expectation matching, giveaway logic, and tone recognition. Flags genuine misrepresentations or important omissions.

Agent 20 / Final Review

Three passes:

1

Slop Scan

16-point check for filler words, vague language, and anything that sounds like it was written by a chatbot.

2

Phase Completeness Check

Confirms every required element made it through the pipeline. No missing hooks, no dropped CTA options, no gaps.

3

Character Count + Style Match

Enforces 1,400-character max and scores against the style library. Assembles the final script with all approved hooks, body, and CTA options.

The Bottom Line

20 agents. 5 phases. 6+ adversarial gates. Default position: REJECT. What survives this system is a script that has been researched against real platform data, war-tested line by line, and verified against the founder's own words. That is why their launches average 2 million views per video.

Keep reading

Get the rest of “20 AI Agents That Write Viral Launch Videos” + the PDF — free.

Drop your email to unlock the full guide and download it.

joshualarosa.ai

Quiet editorial · siblings with building leverage

InstagramTikTokYouTubeSubstack