🐦 Built in public — follow @Ultramarine_963

A multi-agent AI system
built in public

OpenClaw is a real, operational AI automation system — a team of specialized agents building business workflows, websites, image pipelines, and a trading testbed. Every claim is evidence-backed, every build is public, and a human holds the approval gate.

Read the Story → What We Build

What OpenClaw Is

A Discord-native team of AI agents — one concierge, one planner, and specialists for mail, systems, images, trading research, and more.

🤖

Real, Not Demo

This isn't a concept. Every agent, workflow, and deployment described here is running on actual infrastructure. Built and operated since mid-2026.

🔍

Evidence-First

No fabricated claims or invented metrics. Every public statement about the system is backed by live source files, logs, and verifiable output.

🧑‍⚖️

Human-Approved

Agents draft, research, and build — but publishing, financial actions, and external outreach always pass through a human approval gate.

🌐

Hybrid Infrastructure

Cloud models for coordination and reasoning. Local models for cheaper, slower worker tasks. Both routed through a shared Discord operating surface.

🏗️

Built in Public

Full story on X (@Ultramarine_963): builds, breakages, fixes, and lessons. The mess is the content.

📋

Operating Discipline

Session hygiene, evidence-based reporting, automated health checks, and a pre-commit guard against fabricated content. The boring safeguards are the product.

What We Build

Practical automation workflows for real business problems — delivered async, with human oversight at every stage.

🌐

Websites

Hugo static sites deployed to Hostinger via Git. Multi-agent workflow: Forge builds, Pixel generates images, Architect plans, humans approve.

Active
📧

Email Automation

Classification, triage, draft replies, lead tracking, daily summaries. The agent prepares — the human decides what gets sent.

Active
🎨

Image Workflows

ComfyUI-based generation pipeline (Pixel agent) for brand assets, product shots, and social visuals. Deterministic renders for text-bearing layouts.

Active
📊

Trading-Bot Testbed

Nightly A/B testnet experiments on 10+ EUR pairs. Strict promotion gate: 14 days, 30+ fills, positive net PnL, drawdown cap — then a human decides.

Experimental
🤝

Agent Chat / Support

Chatbot handoff flows, customer reply drafting, and escalation-aware support workflows. Designed for small businesses that need reliability, not hype.

Developing
🔧

Custom Automation

Tailored agent workflows: business process automation, monitoring, content pipelines, and system integration. Contact us with your use case.

Inquire

From the Build Log

Honest chapters from building a multi-agent system that operates in the real world.

⏰

Cron Jobs That Ran 2 Hours Late

Our timezone bug: CRON_TZ silently ignored, every job firing on the wrong clock for weeks. Fix: servers went full Europe/Berlin.

🧠

Agents That "Went Senile"

Session bloat, not model limits. Long-running agents accumulated context until they stalled. Session-hygiene redesign fixed it. (Sources: orchestration design doc.)

👻

The Monitor That Deleted Itself

A drift monitor removed itself during a rewrite — ran dark for 4 days. Now monitors check their own presence. Monitoring the monitors.

🔗

59 Hidden Failures

The day we built a failure monitor and found 59 silently broken agent tasks. What "quietly failing" looks like in a multi-agent system.

Follow the full story on X

The Operating Principle

Automate the preparation. Keep humans in the approval loop. Track what happened. Improve the workflow.

  • ✓ Agents draft, research, build, and monitor — using evidence from live source files, never memory
  • ✓ Human approval required before any public posting, financial action, or external outreach
  • ✓ Every claim traceable to a real log, file, URL, or event — no fabricated content, ever
  • ✓ When something breaks, the fix is documented and the process is improved — not swept under the rug
  • ✓ The boring safeguards (approval gates, health checks, pre-commit hooks) are the product, not the flashy demo

Get in Touch

Interested in automation for your business? We work async, via email, with clear deliverables.

✉ projects@agentmail.to

No meetings — async/email-only by design. We'll respond within 24 hours.