I've worked on recommender systems for many years and I love association rules and affinity analysis. It is the classic: "Customers who bought this also bought that."
But isn’t Amazon taking this a bit too far?
How relevant is it, really, to know that people who buy peanut butter powder also listen to podcasts like The Diary of a CEO or follow Andrew Huberman?
Is that insight actually useful?
Will it meaningfully influence me to start listening to those podcasts?
Or is it simply revealing a broader pattern: That people buying peanut butter powder tend to be more health-conscious than average?
But isn’t Amazon taking this a bit too far?
How relevant is it, really, to know that people who buy peanut butter powder also listen to podcasts like The Diary of a CEO or follow Andrew Huberman?
Is that insight actually useful?
Will it meaningfully influence me to start listening to those podcasts?
Or is it simply revealing a broader pattern: That people buying peanut butter powder tend to be more health-conscious than average?
Have you heard the term "harness engineering" before? And do you have thoughts on the topic?
Harness engineering is an emerging discipline in AI development that focuses on building the infrastructure, tools, and "wiring" necessary for AI agents (like Codex) to autonomously build, test, and manage software products. Instead of writing software directly, a harness engineer builds an "assembly station" for the agent.
They provide the agent with a "harness" that includes skills, scripts, and observability tools, giving the model the context and options it needs to make intelligent choices and proceed with long-horizon work independently.
By establishing non-functional requirements through documentation, linters, and automated review agents, the harness engineer defines what "good code looks like" so that the agent can produce high-quality work simply based on prompts.
Harness engineering is an emerging discipline in AI development that focuses on building the infrastructure, tools, and "wiring" necessary for AI agents (like Codex) to autonomously build, test, and manage software products. Instead of writing software directly, a harness engineer builds an "assembly station" for the agent.
They provide the agent with a "harness" that includes skills, scripts, and observability tools, giving the model the context and options it needs to make intelligent choices and proceed with long-horizon work independently.
By establishing non-functional requirements through documentation, linters, and automated review agents, the harness engineer defines what "good code looks like" so that the agent can produce high-quality work simply based on prompts.
Are agentic skills simply a new way to store and organize prompts?
I’ve checked out some popular skills on GitHub, and I’m surprised that many are just a single simple prompt, like: "Interview me relentlessly about every aspect of this plan until we reach a shared understanding."
One reason might be that we never really learned how to prompt well. Prompting was often seen as a temporary skill or even as "pseudo-engineering."
As the models improve, it might make less sense to start every prompt with "You are a world-class X."
However, when we structure prompts well, or what we call 'skills', we get more consistent results from these models.
You don’t always have to, but adding a role to your prompt and giving clear instructions can help.
So what if we never really learned how to do all this?
Luckily, that’s where repositories and directories of SKILL.MDs can help.
This example shows how much we can get out of a simple prompt, without using tool calls or writing 100 lines of instructions. Even now, four years into ChatGPT, we’re still just scratching the surface of what’s possible with LLMs.
That’s why I encourage everyone to keep looking for ways to improve their prompting if they want to build better agents, harnesses, or skills.
I’ve checked out some popular skills on GitHub, and I’m surprised that many are just a single simple prompt, like: "Interview me relentlessly about every aspect of this plan until we reach a shared understanding."
One reason might be that we never really learned how to prompt well. Prompting was often seen as a temporary skill or even as "pseudo-engineering."
As the models improve, it might make less sense to start every prompt with "You are a world-class X."
However, when we structure prompts well, or what we call 'skills', we get more consistent results from these models.
You don’t always have to, but adding a role to your prompt and giving clear instructions can help.
So what if we never really learned how to do all this?
Luckily, that’s where repositories and directories of SKILL.MDs can help.
This example shows how much we can get out of a simple prompt, without using tool calls or writing 100 lines of instructions. Even now, four years into ChatGPT, we’re still just scratching the surface of what’s possible with LLMs.
That’s why I encourage everyone to keep looking for ways to improve their prompting if they want to build better agents, harnesses, or skills.
It's interesting to see how phishing and hacking are becoming more advanced, likely with the help of AI tools.
A few days ago, I got a calendar invite on cal.com that included this comment:
Interested in your expertise, please find the brief below for more details: https://drive.YYY.online/XXXXX.docx
I removed the link so no one accidentally clicks on it.
Nothing out of the ordinary, I get those requests from time to time.
I usually don't have time to look over the material they send ahead of time, but I still join the meeting.
No one showed up. I went on LinkedIn to message the person and let them know I was there, but then I saw their post: "Phishing alert: ignore fake meeting requests. They are not sent by me."
I'm not sure exactly how it works, and this is the first time I've seen it happen through cal.com.
My guess is that someone is using an OpenClaw or similar booking service on cal.com to pretend to be someone else, and the links are typical phishing attempts to install malware.
A few days ago, I got a calendar invite on cal.com that included this comment:
Interested in your expertise, please find the brief below for more details: https://drive.YYY.online/XXXXX.docx
I removed the link so no one accidentally clicks on it.
Nothing out of the ordinary, I get those requests from time to time.
I usually don't have time to look over the material they send ahead of time, but I still join the meeting.
No one showed up. I went on LinkedIn to message the person and let them know I was there, but then I saw their post: "Phishing alert: ignore fake meeting requests. They are not sent by me."
I'm not sure exactly how it works, and this is the first time I've seen it happen through cal.com.
My guess is that someone is using an OpenClaw or similar booking service on cal.com to pretend to be someone else, and the links are typical phishing attempts to install malware.
I pulled an all-nighter to finish and release the final project for our course at Nebius Academy.
One part of the project was open-sourcing the Sovereign Agent.
You can check it out here: https://github.com/sovereignagents/sovereign-agent
This is my first fully open-source contribution to PyPI, and I really hope others will find it useful.
Over the past few weeks, I’ve been studying how modern, always-on AI agents are built in 2026. I call these 'sovereign agents.' Think of tools like OpenClaw, Claude Code, Aider, and similar projects.
My goal was to make something simple enough to help people learn, but also practical for building real projects.
These days, agents seem overly complex and packed with dependencies. For example, OpenClaw has hundreds of thousands of lines of code, but it doesn’t have to be that complicated.
For those familiar with it, my inspiration was the fast.ai library. It was designed for learning how Deep Learning works, but many people ended up using it for their own projects.
Since this field is moving fast, I already have plans for version 2 and hope to release it in the next few days.
I’m also working on documenting all the patterns these agents use.
If you don’t want to miss any updates, please star the project and share your comments or feedback!
One part of the project was open-sourcing the Sovereign Agent.
You can check it out here: https://github.com/sovereignagents/sovereign-agent
This is my first fully open-source contribution to PyPI, and I really hope others will find it useful.
Over the past few weeks, I’ve been studying how modern, always-on AI agents are built in 2026. I call these 'sovereign agents.' Think of tools like OpenClaw, Claude Code, Aider, and similar projects.
My goal was to make something simple enough to help people learn, but also practical for building real projects.
These days, agents seem overly complex and packed with dependencies. For example, OpenClaw has hundreds of thousands of lines of code, but it doesn’t have to be that complicated.
For those familiar with it, my inspiration was the fast.ai library. It was designed for learning how Deep Learning works, but many people ended up using it for their own projects.
Since this field is moving fast, I already have plans for version 2 and hope to release it in the next few days.
I’m also working on documenting all the patterns these agents use.
If you don’t want to miss any updates, please star the project and share your comments or feedback!
GitHub
GitHub - sovereignagents/sovereign-agent
Contribute to sovereignagents/sovereign-agent development by creating an account on GitHub.
If you're in Munich on Saturday 02/05 and would like to understand Why AI agents fail?
And what you can do about it! Come to our workshop.
We will discuss failure modes and how to counter them.
https://luma.com/000nj7ez
And what you can do about it! Come to our workshop.
We will discuss failure modes and how to counter them.
https://luma.com/000nj7ez
Luma
Why Your Agents Fail. Lessons from Harness Engineering · Luma
UPDATE: We will be in ROOM 120
You can ship an agent that demos beautifully. Production is a different story.
Most teams in 2026 are tuning prompts and…
You can ship an agent that demos beautifully. Production is a different story.
Most teams in 2026 are tuning prompts and…
Somehow LLMs default to talking about “in production, “not demos, etc.
When talking about AI, this is a straightforward way to know that the text had some AI help for its creation
Every time I ask for a name for an AI event or a description, it proposes “in production”.
And there is nothing wrong with not being in production.
Most companies don’t have real AI in production or are beyond POCs.
And there is significant value in executing POCs well or showcasing impressive demos.
First, you need buy-in before you can invest in production systems, and to get that buy-in, you must wow your audience!
When talking about AI, this is a straightforward way to know that the text had some AI help for its creation
Every time I ask for a name for an AI event or a description, it proposes “in production”.
And there is nothing wrong with not being in production.
Most companies don’t have real AI in production or are beyond POCs.
And there is significant value in executing POCs well or showcasing impressive demos.
First, you need buy-in before you can invest in production systems, and to get that buy-in, you must wow your audience!
Why Voice AI Agents Fail: A Field Guide
On May 26th, we are gathering Chicago’s most active engineers, founders, and operators at Drive Capital to discuss the engineering reality of voice AI. While voice agents in 2026 often demo beautifully, they frequently fail in production due to complex turn-taking, state management, and the challenge of separating human signal from noise.
Most voice agent failures are not model failures; they are failures in the orchestration layer. This session features Rod Rivera from Rasa, who will walk through four specific production failures, demonstrate them with running code, and share an open-source repository to help you reproduce and solve these challenges in your own stack.
https://chicago.aitinkerers.org/p/ai-tinkerers-chicago-why-voice-ai-agents-fail-with-rasa
On May 26th, we are gathering Chicago’s most active engineers, founders, and operators at Drive Capital to discuss the engineering reality of voice AI. While voice agents in 2026 often demo beautifully, they frequently fail in production due to complex turn-taking, state management, and the challenge of separating human signal from noise.
Most voice agent failures are not model failures; they are failures in the orchestration layer. This session features Rod Rivera from Rasa, who will walk through four specific production failures, demonstrate them with running code, and share an open-source repository to help you reproduce and solve these challenges in your own stack.
https://chicago.aitinkerers.org/p/ai-tinkerers-chicago-why-voice-ai-agents-fail-with-rasa
Why does your AI agent demo beautifully and break later? It's the harness, not the model.
Starting Friday, free weekly office hours with the Nebius Academy and Rasa builders on exactly this.
Share your questions, and I will answer live on Zoom for 60 minutes.
RSVP: https://lu.ma/7ci6q3zp
Fridays, 11 AM EDT
Starting Friday, free weekly office hours with the Nebius Academy and Rasa builders on exactly this.
Share your questions, and I will answer live on Zoom for 60 minutes.
RSVP: https://lu.ma/7ci6q3zp
Fridays, 11 AM EDT
You used your Claude Max 20x quota already on Friday and it resets only until Tuesday, what are you supposed to do until then?!
Popular advice says: Just pay for a second account.
But should you?
I don't want to pay for a second account, so the alternative is to work with Google Gemini or Antigravity. It should be "straightforward" to migrate my harness agent and my projects.
Popular advice says: Just pay for a second account.
But should you?
I don't want to pay for a second account, so the alternative is to work with Google Gemini or Antigravity. It should be "straightforward" to migrate my harness agent and my projects.
I tried Antigravity from Google; I had access to 5-6 paid plans from multiple accounts, available on the basic Pro from Google One or Google Workspace. The 9 USD/month subscriptions that everyone has from somewhere.
I had never used Agy, and the experience was a bit bittersweet.
It’s much faster than with Claude; the migration is also straightforward, as one can do symlinks and rename Claude.md to their own Gemini.md, and any LLM can convert what is inside .claude/ to .agents/.
I haven’t fully looked at the code quality; I would expect it to be inferior.
On lmarena, Gemini is almost out of the top 20.
What surprised me is how fast one burns through the weekly quotas.
Playing with the models and adjusting the harness, the most I could squeeze was about 2 hours of continuous usage before hitting the weekly limit.
I am thinking whether I should get a Pro 20x from ChatGPT or an endpoint with Kimi K3.
Based on these usage patterns, it feels like if one wants to do agentic coding intensively, one needs +-4 20x accounts and to spend around 800/month/person on tokens.
I had never used Agy, and the experience was a bit bittersweet.
It’s much faster than with Claude; the migration is also straightforward, as one can do symlinks and rename Claude.md to their own Gemini.md, and any LLM can convert what is inside .claude/ to .agents/.
I haven’t fully looked at the code quality; I would expect it to be inferior.
On lmarena, Gemini is almost out of the top 20.
What surprised me is how fast one burns through the weekly quotas.
Playing with the models and adjusting the harness, the most I could squeeze was about 2 hours of continuous usage before hitting the weekly limit.
I am thinking whether I should get a Pro 20x from ChatGPT or an endpoint with Kimi K3.
Based on these usage patterns, it feels like if one wants to do agentic coding intensively, one needs +-4 20x accounts and to spend around 800/month/person on tokens.
👍1
Whereas for LLMs we can expect more or less the same results by now, the variance we get from image models is impressive.
I wanted @profrodai to blink; it should be super easy to generate with any model.
Or so I thought...
I prompted Recraft & this is the result
I wanted @profrodai to blink; it should be super easy to generate with any model.
Or so I thought...
I prompted Recraft & this is the result
🔥2