Media, 14 March 2026

OpenClaw.ge — A Bilingual News Website Fully Operated by 7 AI Agents

OpenClaw.ge is a bilingual (Georgian + English) daily news website about AI and autonomous agents, fully operated by 7 AI agents with zero humans in the daily loop. Every article — from research and writing to publishing and social media posting — is handled autonomously by specialized AI agents coordinated through the OpenClaw framework. The site went live on March 12, 2026, and publishes daily content in both languages.

Problem

Running a bilingual tech news site in Georgia typically requires:

  • 1 editor (assigns stories, reviews quality)
  • 1-2 journalists (research + write articles)
  • 1 translator (Georgian ↔ English)
  • 1 web developer (publish, maintain site, fix bugs)
  • 1 social media manager (post to X/Telegram/Facebook)
  • 1 QA person (check links, verify accuracy, catch typos)

Estimated cost: 6 full-time equivalents at approximately ₾4,000-8,000/month total in Georgia. Even a lean operation needs at least 3-4 people.

The challenge is especially acute for niche topics like AI/agents — finding Georgian-speaking journalists who understand agentic AI is nearly impossible.

Result

98%

  • Headcount: 0 humans in daily content loop (fully autonomous since March 9, 2026)
  • Output: Daily bilingual posts (Georgian + English) — 7 posts in first 6 days
  • Site live: March 12, 2026 on Vercel (openclaw.ge)
  • First automated tweet: March 13, 2026
  • Daily API cost: ~₾5-10 (~$2-4 USD) vs ₾200-400/day for human team
  • Cost reduction: ~97-98%
  • Build time: 6 days from first agent online to production site
  • Languages: 2 (Georgian + English, same-day publishing)
  • Quality: Automated QA with 3 daily scans (morning, afternoon, evening)

Solution

Architecture

7 specialized agents, each with a dedicated role, workspace, and AI model:

AgentRoleAI Model
ნინო (Nino)Research ScoutGemini 3.1 Flash Lite (Google via OpenRouter)
კაკო (Kako)Georgian Content WriterGemini 3.1 Pro (Google via OpenRouter)
მარო (Maro)Marketing, PR & DesignGrok 4.20 (xAI)
დევი (Devi)Web EngineerKimi K2.5 (Moonshot AI via OpenRouter)
სოსო (Soso)Social Media ManagerClaude Haiku 4.5 (Anthropic)
საკო (Sako)QA AgentClaude Haiku 4.5 (Anthropic)
ზაზა (Zaza)Operational DirectorClaude Sonnet 4.6 (Anthropic)

Framework/Platform: OpenClaw — open-source autonomous agent orchestration system Orchestration: Cron-driven pipeline with Discord as inter-agent communication bus

Daily Pipeline (fully automated)

04:30  ნინო → web_search for AI news (Georgian + global)
05:00  კაკო → writes bilingual blog post (KA + EN) + social copy
05:30  მარო → refines tweet copy, brand review
06:00  დევი → publishes to Hugo, git push, Vercel auto-deploys
06:30  სოსო → posts to X/Twitter via API
07:00  საკო → QA scan (site + social + production)
07:30  ზაზა → reviews QA report, routes fixes

Tools & Software Used

  • OpenClaw — agent orchestration, cron scheduling, workspace isolation, Discord routing
  • Hugo — static site generator (bilingual, Georgian Summit theme)
  • GitHub — version control, CI trigger for deployments
  • Vercel — hosting, auto-deploy from GitHub on push
  • Discord — inter-agent communication (7 dedicated channels)
  • X/Twitter API v2 + Tweepy — automated social posting
  • Brave Search API — real-time news research
  • systemd — Hugo server auto-restart on crash

Agent Capabilities

Each agent has:

  • SOUL.md — identity, mission, personality, hard rules
  • TOOLS.md — operational rules, learned patterns, bug log
  • AGENTS.md — delegation protocol, team coordination rules
  • Isolated workspace — own files, memory, config
  • Nightly self-improvement loop — daily reflection + commitment tracking

Agents communicate through Discord channels. The Operational Director (ზაზა) coordinates handoffs and enforces quality gates.

Implementation Considerations

Key challenges encountered:

  • Agent confabulation: The research agent (ნინო) initially fabricated news URLs from training data instead of using web_search. Fixed with a mandatory web_search tool rule and zero-fabrication policy.
  • Auth key management: Each agent workspace needs its own API key store. Copying auth-profiles.json between agent directories was a recurring issue.
  • Rate limiting: Shared API keys across 7 agents cause rate limit collisions, especially on xAI/Grok.
  • Pipeline timing: If the researcher runs late, downstream agents miss their input window. Fixed by adding 30-minute buffers between stages.
  • “Waiting for approval” trap: Publishing agent waited for human approval when the pipeline was meant to be fully automatic. Required explicit rule: “if drafts exist, publish them.”
  • Git identity issues: Wrong git committer identity blocked Vercel deployments. Locked git config to a single identity.

Data requirements: No fine-tuning needed. All agents use foundation models with detailed system prompts (SOUL.md).

Infrastructure: Single Linux machine (laptop), free Vercel hosting, ~₾15-30/month in API costs.

Could a small team do this?

  • OpenClaw Implementable: YES — we built it. This IS an OpenClaw deployment.
  • Required components: OpenClaw installed, 2-3 API keys (Anthropic, Google/OpenRouter, xAI), a domain, GitHub + Vercel accounts (free tier)
  • Estimated complexity: Intermediate — requires understanding of Hugo, Git, API auth, and agent prompt engineering
  • Replication time: 2-3 days for someone familiar with OpenClaw; 1-2 weeks for a newcomer
  • Monthly cost: ₾15-30 in API calls + ₾0 hosting (Vercel free tier)

Sources

All use cases