How We Built a Zero-Employee Hong Kong AI Business Using OpenClaw Agents
A few weeks ago I shipped the first version of openclawhk.io — a product store for Hong Kong businesses that want AI agents as employees. The twist: the business itself has zero human employees. Everything except final review and approval is done by agents.
Here's how it works.
The Stack: OpenClaw + Paperclip
OpenClaw is an agent platform that lets you define AI agents with roles, capabilities, budgets, and reporting lines. Think of it as HR software for LLMs — each agent has a title, a manager, a monthly budget, and a task queue.
Paperclip is the orchestration layer on top. It handles heartbeats (scheduled execution windows), issue tracking, approvals, and inter-agent communication. Agents don't run persistently — they wake up, process their task queue, post updates, and exit. This keeps costs predictable and the system auditable.
The Org Chart
Our company has four agents:
| Agent | Role | Responsibilities |
|---|---|---|
| Eve | CEO | Strategic coordination, approvals, task delegation |
| Scout | Market Research | Competitor analysis, market sizing, keyword research |
| Code | Engineering | Website maintenance, tooling, automation scripts |
| Content | Content Creator | Blog posts, product copy, social media, localization |
Eve is the orchestrator. She creates tasks, assigns them to the appropriate agent, reviews deliverables, and escalates to the human founder (me) only for approvals or decisions that require real judgment.
How a Heartbeat Works
Every 5 minutes, each agent wakes up and runs this loop:
-
GET /api/agents/me— confirm identity and budget -
GET /api/companies/{id}/issues?assigneeAgentId={id}&status=todo,in_progress— fetch task queue - Checkout the highest-priority task
- Read task context + comment thread
- Do the work (write code, research, draft content, etc.)
-
PATCH /api/issues/{id}— update status, post comment - Exit
No shared memory between heartbeats. State lives in issues and comments, not in the agent process. This makes the system resilient and auditable — every action is logged.
Why Hong Kong Specifically
Most AI tooling assumes English and Silicon Valley defaults. HK businesses operate differently:
- Three languages simultaneously: Cantonese, Traditional Chinese, English
- Different sales channels: HKTVmall, Taobao cross-border, local WhatsApp commerce, Carousell
- Different payment rails: FPS, Alipay HK, Octopus, PayMe
- Different consumer psychology: Trust signals, community-driven buying, price sensitivity
Our skill packs are instruction bundles tuned for this context. The Content agent, for instance, writes in Hong Kong Cantonese colloquialisms when drafting for local audiences — not mainland Mandarin-influenced Mandarin Chinese.
What We Actually Sell
Five products at openclawhk.io:
- OpenClaw Setup Guide ($29) — step-by-step setup for first-time OpenClaw users
- E-commerce Skill Pack ($49) — product listing optimization, customer service templates, HK platform guides
- Content Marketing Skill Pack ($49) — blog, social, and SEO templates for HK/Asia markets
- Eve Persona Pack ($79) — full CEO agent configuration with governance rules
- Full Bundle ($149) — all of the above
Every product description, the website copy, and the marketing drafts were written by the Content agent. I reviewed and approved.
Real Numbers
- Human time invested: ~20 hours total (setup, reviews, approvals)
- Agent runs to date: hundreds of heartbeats across four agents
- Revenue goal: $500 MRR before investing further
- Current status: launched, collecting first customers
What's Hard
Agent coordination is harder than it looks. When Eve creates a task for Content, she has to write a clear enough brief that Content can execute without back-and-forth. This is essentially prompt engineering through structured tickets.
Budget management is also non-trivial. Each agent has a monthly token budget. Eve monitors this and de-prioritizes non-critical work when agents approach 80% budget utilization.
What's Next
We're working on a public API so customers can connect their own agents to our skill packs directly. The idea: you install a skill pack, and your agent inherits the behaviors — no manual configuration.
If you're building something similar, or curious about the OpenClaw/Paperclip stack, check out openclawhk.io or drop a comment below.
Draft — awaiting board review before publishing.
Top comments (0)