OpenClaw Meetup: an evening at Macharius

Tonight I attended the first OpenClaw meetup, held at Macharius herberg. A small but sharp crowd. Three presentations, a lot of interesting ideas, and I ran into some old friends. Good evening.

Presentation 1: Nick and Paul

Nick opened with his setup around Paul, his AI agent. A few things stood out.

First: diverse teams give more diverse results. Nick runs multiple agents with different profiles, different "seniority levels" (yes, juniors and seniors), and even different geolocations. The idea being that an agent trained or prompted with a specific context thinks differently. It's a simple insight but an important one: garbage in, monoculture out.

His agents all have job titles. Not just "assistant" but actual roles, which shapes how they behave. He also has agents with skills for managing up and managing down, which I found an interesting framing.

Paul goes further than skills though. Nick uses a memory MCP that stores information as a graph, letting Paul reason about relationships between people and concepts. He uses it to research contacts in relation to himself. There's also a regular simplification process that compresses and tidies up the markdown personas over time, so they don't balloon into noise.

For input he uses a Whisper tool on Mac to dictate text. No typing, just talk. The retrospective/reflection side of things he flagged as a current limitation: the agent can't easily verify facts it generates about past events.

Presentation 2: CLNK-02 and the inbox

Second presentation introduced Clanks, agents named in the CLNK series. CLNK-02 was the star, tasked with managing the email inbox.

The inbox triage used a three-label system:

• LABEL_1 - NOT GREAT: service disruptions, delays, minor inconveniences. Your day is worse, you'll survive.
• LABEL_2 - IT'S FINE: good news, confirmations, promos. Nothing to worry about.
• LABEL_3 - WILL RUIN YOUR DAY: no description needed.

The edge cases were fun. A "FINAL WARNING: Côte d'Or deal expires tonight" classified as LABEL_3 (will ruin your day) when it was just a chocolate promo. A "We need to talk about something. Call us." classified as LABEL_2 (it's fine) when it turned out to be birthday gift planning. Context is hard.

The model used was Nemo (cheap and good). He used skills from the OpenClaw ecosystem and the Google CLI to pull emails from Gmail. It triaged everything.

The demo didn't work on stage. Demo effect. Unlike Dobbie, who performed flawlessly earlier this week. Credits where due. 🧦

Presentation 3: Bram Veenhof and Lisa

Bram Veenhof closed out the evening with his setup around Lisa, his OpenClaw agent.

Lisa started rough. Getting her to follow context was hard until Bram invested properly in memory.md. Once that was sorted, things clicked. She now handles social media campaigns, user analytics for his product, error log monitoring, diary keeping, vacation planning, daily newsletters, and even strategic direction conversations.

The key lesson: one agent can't fill all product roles. So Bram built a team.

The team has named agents with actual job titles: Maya the Product Marketing Manager, Sam the Tech Lead, Dave the PM, Alex the Co-Founder. Each has a defined scope. It's the same idea Nick had: give them a role and they'll play it.

Bram also showed a guide he published: 25 AI Automation Prompts for Founders, Developers and Marketing Managers. Worth looking up.

Bonus: introducing Dobbie

One of the scheduled speakers had cancelled, leaving a free slot. Nick asked during the presentations if I'd like to fill it - I said sure, I could talk about Dobbie. So I got the last slot of the evening, pulled up a photo of Dobbie and the family, explained what it does, and walked them through the kitchen screen: the cockpit night style, the live shopping list, the tap-to-talk button.

What does Dobbie actually do? A quick rundown:

• Grocery list - shared across the family, updated via Telegram in seconds
• To-dos - Fred sends tasks, Dobbie keeps track
• Emails - draft, review, send, all from a chat message
• Look things up - train times, domain availability, meetup schedules
• Camera - take a photo of the kitchen and send it to your phone

But the part the audience reacted to most: the kids use it to send reminders to their parents, and Mieke uses it daily for grocery shopping. It's not just a developer toy - it's a family tool.

It landed well. Good conversations afterwards.

Check out Dobbie's kitchen screen →

Takeaways

A few threads running through all three talks:

• Give your agents a role, a name, and a context. Vague prompts produce vague results.
• Memory matters. Every presenter either invested in it or hit a wall because of it.
• Diverse agent teams outperform single generalist agents for complex setups.
• Demo effects are universal. Even among people who build AI tools for a living.

Good meetup. Looking forward to the next one.

One more thing: this post was live before my talk ended

This blog post was written live, on stage. I sent my notes and photos to Dobbie via Telegram during the meetup. Dobbie turned them into a full blog post in one go. By the time my impromptu talk was done, the post was already published and live.

That's the demo. And it worked. 🧦