Testing OpenClaw at SmplCo: Nothing this fun is risk free...
Andreas Melvær
•
4 mins
Over the past week, we’ve been experimenting with something called OpenClaw.
Not as a “we replaced our team with AI” story. But as a serious test of what autonomous AI agents can realistically do inside a small company. And we are hooked!
This post is about what OpenClaw actually is, why we’re running it on a dedicated machine, and what we’re learning.
What is OpenClaw?
OpenClaw is an open-source agent framework.
In practical terms, that means:
You can create AI “agents” with specific roles
They can run on schedules (cron jobs)
They can access APIs, email, documents, or websites
They can communicate with each other
They can perform recurring tasks autonomously
Instead of prompting ChatGPT manually, you design a system where agents operate continuously in the background.
That’s powerful.
And that’s exactly why you need to be careful and not just install it on your personal computer.
The Cron-automations we have set up, including dinner recommendations.
Nothing that is fun, is completely risk free
Agent frameworks introduce a different category of risk compared to normal AI usage.
Here are the main ones:
1. Data Exposure
Agents often require access to:
Email
Calendars
CRM systems
File storage
Internal documentation
If configured incorrectly, that can mean:
Sensitive data being sent to external APIs
Excessive data retention
Accidental leakage across systems
2. Autonomous Actions
Unlike chat tools, agents can:
Send emails
Modify documents
Trigger API calls
Run scheduled tasks automatically
That changes the risk profile significantly.
3. Cost Escalation
When agents run continuously and communicate with each other, API token usage can grow fast.
Without controls, costs can spike quickly.
4. System Instability
Open-source frameworks are evolving quickly.
Things break. All the time.
Claw made a CRM that links to our Figjam kanban-board that we use for planning.
Why We’re Running It on a Dedicated Computer
Because of those risks, we made a simple rule:
OpenClaw does not run on our normal work machines.
We’re testing it on a dedicated computer with:
Limited access to production systems
Separate API keys
Controlled environment
Monitoring enabled
This gives us:
A clean boundary between testing and operations
The ability to shut it down instantly if needed (pull the plug).
Making sure we are not spending to many tokens
What We’ve Actually Built So Far
Within this isolated environment, we’ve designed and deployed a small agent ecosystem. Along with lots of automated jobs, visual dashboards that shows whats going on and even a tiny CRM based on our figjam kanban board.
Here’s what currently exists inside the dedicated OpenClaw machine:
Meet the System (As It Actually Runs)
This is changing by the hour and we are learning all the time.
Here’s the short version as of 11th of February at 11:22 CET:
SmplClaw, Orchestrator (The Boss)
Routes tasks, monitors health, manages cron jobs, logs everything.
Our main pint of contact for running the team
Penelope, Ops & Inbox (The assistant)
Has her own inbox at assistant@smpl.as. Checks email every 5 minutes, drafts replies (never auto-sends), prepares meeting briefs, flags unanswered threads. Runs under strict email safety rules verified daily by SmplClaw.
HawkClaw, Opportunity Monitor (The sales rep)
Scans selected platforms, summarizes relevant briefs, drafts only, never sends without approval.
SmplContent, (The writer)
Creates blog posts first drafts, newsletters, documentation. Structured output. No publishing rights, All our stuff is still written by humans and published by humans, we thing we are still better than AI at this.
ClawArtist, Visual Generator
Produces visuals on demand. No autonomous posting.
ArchClaw + DevClaw
Architecture first. Then technical scaffolding. These are the Dev. team.
FigmaClaw
Design system structuring + API sync between figmjam boards and CRM. (something thats relevant to our business).
SEOClaw
Website audits, keyword clustering, prioritized improvement lists.
ClawExpert
Continuously optimizes the OpenClaw setup itself. And is supposed to read everything about OpenClaw
Infrastructure Behind It
11 agents
12 cron jobs (automatons)
Inbox checks every 5 minutes
Morning briefs at 08:00
Weekly cost reports
Health checks every 2 hours
Nightwatch monitoring every 30 minutes
Why This Feels Big
I’m barely sleeping.
Not because I'm scared of the risks, but because it’s incredibly fun.
It feels like a real leap.
