I get my personalized AI news delivered every morning at 7:30 by four agents on my gaming PC.
Not a news service or a dashboard. Just very readable text, in the exact style I want, from four AI agents running locally on an old PC via OpenClaw, scanning the market and delivering three sharp stories with implications for decision-makers. Straight to my Telegram app.
This post walks through how it works, what the setup looks like in practice, what surprised me, and where OpenClaw stands today if you're considering building something similar.
The workflow
StefAI is the orchestrator. It manages the entire task end-to-end, every single day, without me touching anything.
Every morning it launches a researcher agent to scan the web for the most important genAI news from the last 24 hours. The researcher's job is coverage: find everything relevant, miss nothing important, skip the noise.
Then the validator agent kicks in. The skeptic. The devil's advocate. It reviews everything the researcher found. Any hallucinations? Are the web sources solid? Are the claims supported? Is there context missing? The validator exists for one reason: to make sure nothing thin or misleading gets through to the next step. And it does not even have access to the web. No need for that.
Once the material is validated, the orchestrator passes it to a copywriter agent that formats a maximum of three stories as short narratives. Each story includes implications for the boardroom. The format is consistent: what happened, why it matters, and what a decision-maker should be thinking about. Readable. Actionable. No filler. I have been tweaking the copywriter system prompt, since I hate it when it sounds like a machine. Now it starts writing like me.
Finally, the orchestrator itself reviews and quality-checks the full output before everything lands in my Telegram channel.
All four agents run on ChatGPT 5.2. Simple. Effective. Cheap.
It's still a pilot. But the quality of the output is already usable. I read the output almost every morning and it consistently surfaces stories I would have found myself, plus a few I would have missed.
Why a separate gaming PC?
This is worth addressing directly. At the time I write this, there are 900+ malicious plugins in the ClawHub repository, representing roughly 20% of the ecosystem and an unknown number of open security issues. So you'd better keep the framework on separate hardware. Keep it far away from your work PC/Mac.
My gaming PC runs the agent swarm. My work machine stays clean. It's a simple precaution, and until OpenClaw closes that security gap, I'd recommend the same to anyone setting this up.
What surprised me
The agent network sticks to the task. No surprises. No problems.
Sounds boring. That's exactly the point.
If you've spent time working with AI agents, you know that "no surprises" is the best quality signal you can get. It means the roles are clear, the permissions are tight enough, and the orchestrator actually orchestrates. Agents going off-script, hallucinating context, or producing inconsistent output is the norm in many setups. This one just works.
I attribute that to two things. First, the role separation. Researcher, validator, and copywriter are three fundamentally different ways of thinking. When you force one agent to do all three, quality drops. When each agent has one job and clear boundaries, the output is dramatically better.
Second, the orchestrator as quality gate. StefAI doesn't just delegate. It reviews. That final check before delivery catches the occasional weak story and keeps the standard consistent.
The practical setup
The tech stack is straightforward:
Win11 with Ubuntu terminal. OpenClaw to run the agent swarm. A cron job to trigger the daily scan. Telegram as the delivery channel.
And one more piece that turned out to be essential: a Claude Opus 4.6 in a separate Project as a technical assistant, with a system prompt configured as an OpenClaw expert. Whenever StefAI can't solve something, or whenever I hit a configuration issue I can't figure out, I switch to Claude and get help.
This pairing turned out to be the real accelerator. OpenClaw documentation not always complete? Problems come up? Having a dedicated AI assistant that knows the framework inside out means you're never stuck for long. Remember to load the relevant links and files from OpenClaw into the Claude Opus project!
Persistence and patience are useful traits when setting this up. There are moments where it feels like nothing works. But almost anyone at "IT superuser" level can do this. The system is logical. The problems are solvable. And with a good technical sparring partner next to you, the friction becomes manageable.
An honest assessment of OpenClaw
OpenClaw is not ready for enterprise. Governance, audit trails, role management at scale, compliance frameworks. None of that is there yet. If you're in a large organization evaluating this for a team rollout, wait.
But for solo entrepreneurs and small startups who want to build their own agent workflows, it's a different story. It's accessible, affordable, and functional today. You can build a working agent network on your own separate machine, with good control over what happens, at a cost that's trivial compared to any SaaS alternative.
The tradeoff is clear: you get power and flexibility, but you take on the security responsibility yourself. Keep it on separate hardware. Be deliberate about permissions. And don't connect it to anything you can't afford to expose.
What I'd build next
The four-agent setup handles my daily market intelligence. But the architecture is extensible. I'm already thinking about what a fifth agent could do. Or if I should repurpose the setup to totally different use cases: searching for the next vacation in Sardinia? Searching and selecting candidates for a new Snowboard? (just sold my old one, and need a new).
The cool part is that whatever the bots do, they build a personal knowledge base locally (all deliverables from OpenClaw agents sit on your machine in well organized folders).
What I will do next? Check out the limits of Claude Code and Claude Cowork for similar use cases. I will write about it next time.
Final reflection
The technology to build personal agent workflows is available today. Not in a year. Today, on hardware you probably already own and it is sitting there.
What would you set AI agents to do for you?
Stefano Vincenti · AI Built Human · aitrainer.dk
Appendix: OpenClaw System Prompts – Collection & Guide
Want to get started with OpenClaw yourself? Download the complete collection of system prompts for the four AI agents (The Project Manager, Researcher, Validator, and Copywriter) as a PDF. The guide includes ready-to-use prompts, a mini installation guide, and tips for adapting the agents to your own workflows.