Skip to content
fethiverseHire me

02Multi-agent content system

Fethiverse

Ten AI agents that research, write, design and publish a daily tech-news post, then learn from how it performed.

Role
Architecture and build
Built
February to August 2026
Runs
Twice a day, on a schedule
Agents
Ten, each with its own identity

In short

Fethiverse is an autonomous Instagram content factory built on the Claude Agent SDK. Specialist agents with their own identities find the day's AI news, plan a Turkish carousel, write it, source the images, render the slides, review them against a scored rubric and publish, twice a day. What each post earns feeds the next run's choices.

A personal media brand needs a good post every day, and a single prompt writes generic ones.

One-shot prompting repeats hooks that stopped working, invents facts and forgets what the last run learned. Fethiverse splits the work across specialists with persistent identities, a shared knowledge store and messages that survive between runs, and closes the loop from engagement back into the next plan.

What it does

  • Finds the news

    A news researcher sweeps Firecrawl, X, 14 RSS feeds, Hacker News and Reddit, merges and de-duplicates the stories, and picks the day's topic.

  • Writes, designs, renders

    A content director plans the carousel, a copywriter and an art director work in parallel, and slides render in a fixed dark editorial look, in Turkish, with no emoji.

  • Reviews before it ships

    An editor agent scores every carousel out of ten and sends each issue to the agent that must fix it. Nothing ships until it approves, within a bounded number of rounds.

  • Learns from results

    About a day after posting, engagement is measured, patterns are extracted, and Thompson sampling picks the next hook style: proven ones more often, new ones still tried.

  • Remixes what's trending

    A viral scout spots breakout posts, a video researcher finds the official source video, and the pipeline turns it into a branded video slide.

Designed and built it

From the first LangGraph prototype to the Claude Agent SDK pipeline that runs today: agents, tools, knowledge store, learning loop, tests and scheduling.

Python decides the order of the steps; each agent runs in its own SDK session with its identity and a briefing. Agents talk in two ways: notes inside a run, and persistent messages that the recipient reads in its next run.

The factory at a glance

Who does what, and where the memory lives. Hover or tab through the parts.

Production agents, one SDK session eachAround the runbriefings in, peer messages outtool callsmetrics after about 24 hremix rounds, analysisone step at a timepublishArt directorCompetitor analystCaption writerContent directorCopywriterData scientistEditorInstagramKnowledge storeLearning cycleNews researcherOrchestrator: limits, alerts,engagementPipeline: Python decides theorderScheduler, twice a dayAgent tools: research,images, render, memory,publishVideo researcherViral scout
What each part does
Art director
Finds or generates images; no fabricated faces.
Competitor analyst
Keeps track of what comparable accounts are publishing.
Caption writer
Writes the caption and hashtags.
Content director
Plans the carousel, using past lessons and the hook styles that are working.
Copywriter
Writes the copy for every slide.
Data scientist
Runs analytics after a run; never blocks publishing.
Editor
The quality gate: scores the slides out of ten and approves or asks for fixes.
Instagram
Carousels are published through the Instagram Graph API.
Knowledge store
A SQLite store with full-text search: lessons, directives, engagement patterns and messages between agents.
Learning cycle
Turns engagement data into patterns and updated strategies for each agent.
News researcher
Scans the day's AI and tech news and picks the topic.
Orchestrator: limits, alerts, engagement
The orchestrator enforces rate limits, sends alerts and tracks engagement.
Pipeline: Python decides the order
Plain Python code decides which agent works next, so the run is predictable.
Scheduler, twice a day
A scheduled task starts the factory twice a day.
Agent tools: research, images, render, memory, publish
55 in-process tools: research, images, rendering, memory and publishing.
Video researcher
Verifies and records the official source video for a remix.
Viral scout
Scores breakout posts from watched accounts for remix rounds.

One run, message by message

A real run shape, including one revision round from the editor. Step through it or scroll.

par · In parallelFind today's topic1News research2Briefing for each agent3Learnings, directives, unread peermessages4Topic, briefing, hook-style odds5Carousel plan and handoff notes6The plan7Slide copy8The plan9Photos and images10Render the slides11Slide images12Review the slides13REVISE, with a score and issues peragent14Keep the Editor's feedback for nexttime15Revision round with the Editor's notes16Fixed copy17Render again18Review again19APPROVE20Approved slides21Caption and hashtags22Save the run and each agent'sreflections23PipelineNews researcherKnowledge storeContent directorCopywriterArt directorRendererEditorCaption writer
All 23 steps as text
What each part does
Pipeline
The Python pipeline that calls each agent in turn.
News researcher
News researcher agent.
Knowledge store
The shared knowledge store.
Content director
Content director agent.
Copywriter
Copywriter agent.
Art director
Art director agent.
Renderer
The renderer: plain code that draws the slides.
Editor
Editor agent, the last quality gate.
Caption writer
Caption writer agent.

How it learns between runs

What happens after a post goes out, and how it changes the next one.

About 24 hours laterPublish the carousel1Read the post's metrics2Store the engagement3Update the patterns for each hook style4Wins and losses per hook style5A sampled hook style for the next run6Insights in the next briefing7OrchestratorInstagramEngagement trackerPattern extractorKnowledge storeHook samplerContent director
All 7 steps as text
What each part does
Orchestrator
The orchestrator that publishes and tracks.
Instagram
Instagram, where the post lives.
Engagement tracker
Reads each post's metrics a day later.
Pattern extractor
Finds which hook styles perform.
Knowledge store
The shared knowledge store.
Hook sampler
Thompson sampling over hook styles: mostly proven ones, some exploration.
Content director
The content director in the next run.

Decisions and trade-offs

  1. From a LangGraph graph to the Claude Agent SDK

    The pipeline moved to the Claude Agent SDK in March 2026, one session per agent with its own identity file; the old graph code was removed and stays recoverable from history.

  2. Two channels for agents to talk

    Handoff and feedback notes inside a run, plus persistent peer messages with a channel, a priority and an expiry. Both go into the user prompt, never the system prompt, so each agent's identity stays clean.

  3. A tool isn't wired until it's called

    A tool can be registered, allowed, defined and described and still never used, because the agent's identity names only its original tools. Verification is a live, hook-instrumented probe that shows the call happened.

  4. Honest about parallelism

    The researcher is told to query five sources at once, but a probe measured them running one after another. The project does not claim parallelism until a real fan-out tool exists.

  5. Three tiers for every new capability

    Memory, personality and multi-turn reasoning become an agent; one structured model call becomes a script; anything deterministic stays plain code.

  6. No fabricated faces

    Images are identity-preserving, with no fabricated faces; when a photo can't be used, the slide falls back to a scene image.

Results

  • 10specialist agentseach in its own SDK session
  • 55agent toolsacross 14 tool modules
  • 2,263test functionsin 110 test files
  • 357commitsFebruary to August 2026
  • 2×a dayscheduled, autonomous publishing

Stack

Agents
  • Claude Agent SDK
  • In-process MCP tools
  • Python
Memory
  • SQLite with FTS5 full-text search
Media
  • Gemini image model
  • Pillow
  • CairoSVG
  • yt-dlp
  • ffmpeg
Publishing
  • Instagram Graph API
  • Telegram run summaries
  • Scheduled tasks