OpenClaw Seminar

AI by Hand ✍️ Seminar Series

You are standing on the rail track. An AI hype train is fast approaching—blaring its horn, high beams blinding. The train’s name is OpenClaw. You see many people trying to jump on. Some get hit head-on. Some are run over. Some barely survive the leap, only to be thrown off moments later.

“What should I do?” you wonder.

In last week’s seminar, I showed a simple way to face this hype train: just take a step to the side, away from the track, watch, and learn.

That small step made a world of difference. We were able to watch every car as it passed, starting with the engine—the Transformer model published in 2017. We studied a total of twelve cars:

1. Transformer

2. Completion

3. Instruction

4. RAG

5. Chat

6. Agent: Tool Use

7. Agent: Memory

8. Agent: Chain of Thought

9. OpenClaw: Heartbeat

10. OpenClaw: Memory

11. OpenClaw: Tool Use

12. OpenClaw: Message Queue

The train was indeed an engineering marvel. It had all the luxuries you’d expect.You saw passengers dining in the dinner car, sipping wine, cheering each other. For a moment, you envied them.

Then, on top of the dining car, you noticed some suspicious figures. Bandits. Each carried an “injector” gun and an abundant supply of “prompt” bullets.

The day after, you saw the news. The train had been robbed. Bullet shells everywhere. On one of them, a transcription: “ignore the system prompt.”

Feedback

Industry Expert: Val Andrei Fajardo

I’m glad to be joined by my good friend, ​Val Andrei Fajardo, author of Build a Multi-Agent System from Scratch—already a Manning early-access bestseller—and a founding engineer of LlamaIndex, where he helped scale the open-source library to tens of millions of monthly downloads and shape a core layer of today’s LLM application stack.

I got to know Andrei when he was building LlamaIndex. I wanted Andrei to share his story: a strong math and statistics foundation, real-world experience deploying AI systems in startups, and visible open work. He helped maintain the rapidly growing library in the open with the community, learning not just how to design features, but how to choose the right abstractions and integration boundaries so a system can scale.

One audience member asked an interesting question: “What keeps you awake at night?” I jumped in and said I already knew the answer—it had to be his newborn, his third boy—which got a smile. But Andrei then gave a very serious answer: he’s concerned about the future of software built from unchecked coding-assistant output, and the fragile systems and poor user experiences that could follow. That concern is exactly why he teaches people to build agents from the ground up, not just assemble them.

Andrei started working on Build a Multi-Agent System from Scratch back in early 2025. I recall vividly a few months later (I believe it was July), I asked Andrei how his new book was coming along. He said “The book writing is a slog. I do find it similar in ways to writing a PhD thesis.”

slog (noun / verb): a period of long, exhausting, repetitive work that requires persistence but feels slow and heavy.

This is just one of the many amazing diagrams Andrei created through his slog, a sharp contrast with weekend vibe-coding projects like OpenClaw. I asked Andrei to do a live walkthrough of this diagram during the seminar. He graciously agreed. He walked us through the full tool-calling lifecycle: the LLM produces a tool request in text, a separate runtime executes it, and the result is fed back into the conversation for synthesis. This becomes a complete framework—runtime, processing loop, tools, memory, and model integrations—culminating in a local personal assistant, an OpenClaw-style system built entirely from first principles.

Recording & Workbook

(limited time preview)

The full recording and the associated Excel workbook are available to AI by Hand Academy members. You can become a member via a paid Substack subscription.

🔗 Link to the Excel Workbook