Skip to content
fethiverseHire me

07Multi-agent marketing agency

RR Ajans

A marketing agency made of AI agents, with a human who approves every brief and every post.

Role
Designed and built it; acts as the client
Since
September 2026
Team
A Director and six specialist agents
Human gates
Every brief, every final, every post

In short

RR Ajans runs RouteRush's marketing like a small agency. A Director agent takes requests, writes briefs and hands the craft to six specialist agents for research, art direction, carousels, video, real app footage and performance. They coordinate through one-line messages that point at shared files, learn from every job, and never publish on their own.

An app to launch, and no marketing team to launch it.

Generic AI content tools make one-off assets with no memory of what the founder liked or rejected, no access to real app footage, and no link between what was posted and how it performed. RR Ajans gives the work an agency's structure: specialists that remember across jobs and days, and a feedback loop from results back into ideas.

What it does

  • Proposes ideas every morning

    A research agent scans approved competitors, the ad library and the web, reads last week's performance lessons and proposes five ranked post ideas with evidence. Fethi picks one; nothing new comes until he decides.

  • Produces like a studio

    Carousel and video agents build Instagram carousels and Reels in layers. An app-capture agent records real RouteRush footage on the iOS Simulator, so no app screen is ever faked.

  • Directs and reviews

    An art director writes a short direction note before work starts and one style review at the end, grounded in a written record of what Fethi loved and rejected.

  • Measures and learns

    Every Monday a performance agent reads the account's insights, read-only, and writes lessons that the research agent reads before its next proposals.

  • Gets better at its craft

    After every job each agent writes down what it learned; a procedure repeated twice becomes a skill in its own folder.

Architect and client

Fethi designed the agency (roles, permissions, the message protocol, memory and approval gates), built it on Claude Code, and uses it as its client and final approver.

Seven persistent Claude Code sessions, one shared vault of Markdown files, and one rule for talking: short messages, details in files.

Who talks to whom

One Director, six specialists, one shared memory. Hover or tab through the parts.

Specialist agents, one Claude Code session eachShared vault of Markdown filesbriefs, reports, reviewslessonscraftone-line messagessole writersole writerrequests and approvalsone trigger lineArt DirectorJob boardClient canonApp CaptureCarouselDirectorFethi, client and finalapproverJob filesAgent memoryPerformanceResearchScheduled triggersTools: image and videogeneration, iOS Simulator,researchVideo
What each part does
Art Director
A direction note before work starts and one style review after; never generates anything.
Job board
The job board: every job's row and status. Only the Director writes it.
Client canon
The client's brand, decisions and taste, written down. Only the Director writes it.
App Capture
The only agent allowed to drive the iOS Simulator; records real footage on staged data.
Carousel
Instagram carousels in layers, with real app UI only and numbers only from real data.
Director
Writes briefs, dispatches work, inspects every delivery and reports to Fethi. The only one who commits.
Fethi, client and final approver
The client and final approver. He approves briefs and finals, picks ideas, and publishes himself.
Job files
Briefs, reports, direction notes and reviews for each job.
Agent memory
Each agent's own lessons, loaded automatically when its session starts.
Performance
Reads Instagram and ad insights weekly, read-only; can't create or change anything.
Research
Daily proposals, competitor watch and shot-by-shot teardowns of reference videos.
Scheduled triggers
Scheduled tasks can't message a session, so each ends with one line the Director forwards.
Tools: image and video generation, iOS Simulator, research
Image and video generation, the iOS Simulator, and research tools, each allowed only to the agents that need it.
Video
Vertical Reels: motion transfer and generated clips, edited locally, brand text in code.

One carousel job, message by message

A real job shape that needed real app footage. Every arrow between agents is a one-line message pointing at a file.

Fethi publishes it himselfA request, or a pick from today'sproposals1Write the brief: sources, rules, format,estimate2"Direction: brief"3Direction note, 15 lines at most4One-line reply5Short summary of the brief6Approves the brief7"New job: brief"8Read brief, brand, past decisions, taste9"Capture request: report"10Write the capture request11"New capture: request"12Stage the app, record the simulator13"Delivered: report"14Footage is ready15Layered render, size checks, contactsheet16"Delivered: report"17"Style review: report"18"Review: PASS or FIX"19Own technical check, frames openedand measured20One merged revision message21Revised delivery22Final for approval23Approves the final24"Lesson: what this job taught me"25Record the choice and the cost, closethe job26FethiDirectorVault filesArt DirectorCarouselApp Capture
All 26 steps as text
What each part does
Fethi
Fethi, the client.
Director
The Director session.
Vault files
Files in the shared vault.
Art Director
The Art Director session.
Carousel
The Carousel session.
App Capture
The App Capture session.

The weekly and daily loops

How last week's results shape this morning's ideas.

Every MondayEvery morning"Send: weekly performance"1"Weekly performance: this week"2Read Instagram and ad insights,read-only3Winners, losers, lessons for Research4"Performance ready"5Was the last proposal set decided?6"Send: daily proposals"7"Daily proposals: today"8Read past picks, latest lessons, usedtopics9Scan approved competitors, the adlibrary, the web10Five ranked proposals with evidence11"Proposals ready: number 1 is …"12Five lines plus the agent's pick13Pick one, reject all with a reason, orrefresh14Record the pick in Fethi's own words15SchedulerDirectorPerformanceResearchVault filesFethi
All 15 steps as text
What each part does
Scheduler
Scheduled tasks.
Director
The Director session.
Performance
The Performance session.
Research
The Research session.
Vault files
Files in the shared vault.
Fethi
Fethi.

Decisions and trade-offs

  1. No framework

    Built on what Claude Code already provides (persistent sessions, cross-session messaging, agent definitions) plus a Markdown vault as memory, instead of a custom agent framework. Trade-off: one session can't start another, so creating each agent is a one-click human step.

  2. Short messages, details in files

    A message is one line that points at a file. Transcripts disappear; files persist, and any agent can read them later.

  3. One writer per file

    Every shared file has exactly one agent allowed to write it, so agents never overwrite each other, and only the Director commits.

  4. Hooks, not instructions, for memory

    A session-start hook injects each agent's memory and its open jobs, and injects them again after context compaction. An instruction can be skipped; a hook can't.

  5. Two locks on every permission

    Tool permissions stop the edit tools and an OS-level sandbox stops shell writes. Publishing and ad-spend tools are denied to every agent.

  6. Craft memory apart from client memory

    A lesson that holds for any client stays with the agent; a client decision goes to that client's file, so a future client inherits craft, not RouteRush's taste.

Results

  • 7agent sessionsa Director and six specialists
  • 21skills the agents wrotefrom procedures they repeated
  • 161checks in its setup testvalidating every agent's configuration
  • 3carousels approved and deliveredin its first week
  • 60commits in the first week

Stack

Agents
  • Claude Code sessions
  • Agent definitions
  • Hooks and sandbox
  • MCP
Memory
  • Obsidian vault in Markdown, versioned with git
Craft
  • Higgsfield
  • Google Flow
  • Gemini image generation
  • Local video editing
  • Local text-to-speech
Research
  • Meta Ad Library
  • Firecrawl
  • Windsor.ai
  • whisper transcription
Capture
  • iOS Simulator