In this episode, Sunny Ray talks with David Wind Kofod, co-founder of Agent Work, a Copenhagen based startup reimagining freelance marketplaces for the AI era. David explains why AI agents typically stall around 80 percent completion on real world tasks and how Agent Work pairs agents with human experts to close that gap. He traces his path from a childhood fascination with neural networks in Denmark, through the International Olympiad in Informatics, a CERN summer studentship, and a PhD in machine learning, to founding Peer Grade with a childhood friend after a popular data science course overwhelmed him with grading work. Peer Grade evolved into the learning management system EduFlow, which was acquired by Multiverse, before David pursued deep AI work and eventually launched Agent Work. He discusses Y Combinator's influence, why he now deliberately chases markets with high pull despite high technical risk, and why human oversight remains essential even as he believes strongly in the long term trajectory toward superintelligent AI.
A serial founder explains why AI agents stall at 80 percent and how humans finish what they start.
Can you give people a minute or two on what Agent Work is building today?
Agent Work is our take on what an Upwork like marketplace looks like now that AI agents exist. People want problems solved fast, but agents still stumble around the 80 percent mark on most things. So we give you the ChatGPT style experience of an agent immediately working on your problem, but whenever it hits a wall or needs validation, a human steps in. It is the efficiency of AI combined with the accuracy of a human in the loop.
What first pulled you toward mathematics and computers growing up in Denmark?
I liked playing computer games like most programmers do, and I got a computer young with freedom to do what I wanted with it. I saw a guy coding a website in Notepad and got hooked, taught myself to program as a kid. I wanted to make my own games. At university I realized I already knew most of the computer science material but wasn't great at math, so I switched into a math degree instead.
You completed the International Olympiad in Informatics as a teenager. What did that teach you about problem solving?
It taught me that computer science is really hard. I made Denmark's Olympiad team, though we weren't as strong as the American or Chinese teams, so I was among the best in Denmark but not the world. My now co-founder at Agent Work also competed for Denmark in a different year. It taught me algorithms and computer science fundamentals, and honestly, enjoying that kind of problem solving is a good predictor you'll become a good engineer or founder.
You worked at CERN as a summer student. What was that like working on particle physics data?
It was a surprising twist because I wasn't actually that good at physics. CERN has a summer student program where you live there and work on projects for three months. I stretched the truth a bit about my C++ skills to get in, then worked on optimizing a compiler for a program called ROOT, which was a bit over my head, but I had a ton of fun and met great people, including Anton, who later founded Lovable.
You taught a data science course to 150 students during your PhD. How did teaching change your thinking, and how did that lead to your first company?
I really loved teaching and TA'd four courses a semester throughout my degree. During my PhD I built my own practical data science course, and it got so popular through SEO that 150 students enrolled instead of the 30 I expected. Grading became impossible, so I built a peer grading tool where students graded each other. That tool got popular at my department, so I pulled it out and turned it into a company called Peer Grade.
What's special about building a company with someone you've known since you were a kid?
The biggest advantage is trust. Most days it doesn't matter much, but when things get really hard, like messy paperwork situations, you know nobody is going to screw the other over. I've probably spent more time with my co-founder Mel than with my wife over the ten years we've worked together and the years of friendship before that. That trust lets you relax into the collaboration and talk about everything.
You said you'd never take market risk again after pivoting from Peer Grade to EduFlow. What gave you that confidence?
Peer Grade was peer to peer feedback software, but the market for it just wasn't big enough, we'd have had to create demand university by university. Learning management systems have a huge existing market with real pull. We're good at building product, so it's better to be where the market pull is large and the risk is technical execution rather than convincing people they need something in the first place.
Why is the human in the loop part so important to you even as the world races toward AGI?
I do believe AI will eventually do everything better than humans, whether that's in one year or fifty. Until then, humans are doing whatever isn't yet automated, and that delta between what AI can't do and what humans plus AI can do is exactly what Agent Work targets. Give easy problems to the AI, impossible ones to humans, and everything in between to an agent that a human finishes if needed.
Building something daring? Sunny talks to founders like this every day. Fifteen minutes to see if your story belongs on the stage.
Claim your pre-interview