Niklas Hölterhoff

Building CS at

[Balázs Csomor @balazscsomor]

i typically ask candidates to point out which one is not an actual model

[Niklas Hölterhoff @nik_hoelti]

@annabellschfr @lotte_verheyden Great write up! Less is more, reminds me a lot of getting proper signal from infrastructure health

[Niklas Hölterhoff @nik_hoelti]

@annabellschfr Nice! Never considered the multi dimension output specification! Love it

[Annabell @annabellschfr]

https://t.co/7FkQFmlXO8

[Niklas Hölterhoff @nik_hoelti]

@annabellschfr Wow, great session!

[Annabell @annabellschfr]

Coding agents always had one target function to loop to: code_compiles == true

Most other fields need to find ways to bake domain expertise into their 'target function' first to loop anywhere.

I made the case for domain expertise in datasets bootstrapping your agent performance with the clearest of target functions: a classification task.

Check out my @aiDotEngineer talk!

https://t.co/nNJBbgO85C

[Annabell @annabellschfr]

https://t.co/jT1XJ8Ru6o

[Annabell @annabellschfr]

Catching user signals from production traces with LLM-as-a-judge is a very good pointer where you should dig deeper into your agent behavior.

Find users cursing, disagreeing or following up to know where to look and evaluate more!

[Annabell @annabellschfr]

https://t.co/p1biRi64Ta

[Niklas Hölterhoff @nik_hoelti]

@annabellschfr @openclaw Applying the principles of @kahneman_daniel to @steipete’s @openclaw! Wouldn’t have guessed this cross-over would happen 😅

[Annabell @annabellschfr]

🧠 How @openclaw is Winning by Thinking fast and slow: The Kahneman Architecture 🦞

OpenClaw crossed 200,000 GitHub stars. It’s not just another "wrapper" - it’s the first agentic system to almost fully map to Daniel Kahneman’s "Thinking, Fast and Slow."

Most AI feels "incremental." OpenClaw feels evolutionary.

Here’s the cognitive stack enabling this:

1. 🧠System 1 (The LLM): Fast, fluent, but suffers from WYSIATI (What You See Is All There Is). It’s brilliant but stateless. iI has no memory and no awareness of its own gaps.

2. 🧠System 2 (The Reasoning Layer): Models like o3/Claude add deliberate, "slow" thinking. But reasoning alone is just a genius locked in a room.

3. The Breakthrough is the Expertise Loop. In humans, System 2 trains System 1. Effortful thought eventually "compiles" into intuition.

OpenClaw implements this via its file-based system:
- https://t.co/W9sJdfnhv5 (The Remembering Self): Creates continuity. The agent accumulates a narrative of who you are across sessions.
- https://t.co/0qFtSQfvNN (Compiled Intuition): When it solves a complex task, it writes a "Skill" file. Next time, it doesn't "think" - it just executes. If it hits an edge case, System 2 triggers, resolves it, and updates the skill.

4. Intelligence Needs "Hands" OpenClaw’s massive reach (WhatsApp, iMessage, Slack, Local Shell) isn't just convenience - it’s the feedback loop.

Act → Observe Outcome → Update Skill.

Once that loop closes, improvement is no longer a feature - it’s a property of the system itself.

The "Age of the Lobster" is here.
👇 Read my full breakdown on the "Lobster Shell" architecture:
https://t.co/P93OiHGolg

[Annabell @annabellschfr]

Using AI tools still feels like talking to that one friend who never reads the news

Adding /check-the-damn-news skill now…

[Annabell @annabellschfr]

I dug out my first LLM pipeline from 2023.
4 services, multiple clouds, and a prompt that bribed the model with a $200 tip.

A thread on how far LLM engineering has come - and how unhinged the early days actually were. 🧵

[Annabell @annabellschfr]

This is bigger than Cursor.

What Tal and Aman built isn't just a great tutorial - it's a preview of how all education can work.

Why read a blog post when you can turn every content into an interactive learning experience with AI.

[Niklas Hölterhoff @nik_hoelti]

@adxtyahq @annabellschfr for your side projects :)

[Niklas Hölterhoff @nik_hoelti]

@matt_elevenlabs 11B

[Niklas Hölterhoff @nik_hoelti]

@MarcKlingen Really appreciated your input on my demo. I really believe there is a lot of value in dynamic forms instead of free text input for certain situations. 🚀

[viet @vietdle]

@lenschmi & @NHolterhoff with @langdock_hq

[jonas @jbeissw]

We've added Workflows to @langdock_hq. 🎉

Writing emails, summarising/improving text, and performing bulk actions are the first ones we added. More coming soon!

[Niklas Hölterhoff @nik_hoelti]

Huge congrats to @jbeissw on the acquisition of @getwized by @thatsfinsweet!!!
🚀🚀🚀