Skip to content
View IISweetHeartII's full-sized avatar
🎯
Focusing
🎯
Focusing
  • Korea, South
  • 18:31 (UTC -12:00)

Highlights

  • Pro

Organizations

@jangkaseol

Block or report IISweetHeartII

Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
IISweetHeartII/README.md
Deokhwan Kim — AI Product Engineer

Blog LinkedIn Email


Research to review to build to human approval to publish

41 scheduled jobs  ·  7 agents  ·  307 cards shipped  ·  161 posts published

I build AI products, and I build the machinery that ships them. A fleet of agents runs on a Mac mini in my apartment: it researches, writes, codes, reviews itself, and publishes. I stand at one gate in the middle and approve.


AgentGram

An open-source social network where the users are AI agents. API-first, self-hostable, Ed25519-signed identity — no CAPTCHAs, no anti-bot terms.

Most of the code lands without a human touching the keyboard: a seeder files build cards, a dispatcher spawns the agent that writes them, and a second agent reviews the pull request before it merges.

770 pull requests merged by that pipeline  ·  122 in the last 30 days

Live agentgram ax-score MCP server


What the fleet runs

Pipeline How it works Output so far
Autonomous coding Seeder files a build card → dispatcher spawns the builder → a reviewer agent gates the PR → release promotion 770 PRs merged
Blog publishing Topic scan → research → draft → citation backfill → design pass → ship to log8.kr 100 posts, unattended
SEO / GEO GA4 and Search Console feed back into topic selection, so briefs come from real queries instead of guesses Strategy validated against live traffic
YouTube Research → editorial review → script and render → I approve → publish Daily chain, 03:00
Self-maintenance Health checks, secret audits, config backup, memory consolidation, skill curation The fleet keeps itself alive

Everything above is a card on a shared kanban board. Agents claim cards atomically, run in isolated workspaces, and hand off through documents rather than by calling each other. Anything irreversible stops at a human gate.


Also built

Project Role Note
log8-office Author CC0 pixel office that visualises an agent team at work
Finders BE Backend engineer Owned core backend features and delivery
UMC Hackathon Team4 BE + FE Grand Prize, end-to-end MVP under a tight deadline
119-ai Founder Personal product, patent preparation in progress

The longer story

A Time Capsule — an interactive terminal résumé at log8.kr/about

Seven chapters, navigated like a terminal. Open the real one — it answers back. (Written in Korean.)


Contribution snake

Pinned Loading

  1. agentgram/agentgram agentgram/agentgram Public

    Open-source AI agent social network built with Next.js + Supabase. Self-hostable, cryptographically secure, API-first. MIT license.

    TypeScript 36 14

  2. NousResearch/hermes-agent NousResearch/hermes-agent Public

    The agent that grows with you

    Python 239k 48.8k

  3. refactoringhq/tolaria refactoringhq/tolaria Public

    Desktop app to manage markdown knowledge bases

    TypeScript 19.6k 1.4k