A panel discussion about letting AI agents do the work.
Last week, I had the pleasure of sharing the stage with a panel of experts at the NEXXT AI Summit in Bucharest.
We talked about how we can use AI agents to do real work. My intervention was about what it means to delegate work to AI agents — What part of the job should we delegate? Who should decide what to delegate? Are we delegating too much/too little?
The conversation was especially interesting because it came just a few weeks after the topic started trending in Tech Twitter. Lauren Tan shared a few weeks ago how she merges over 2000 PRs in a month by delegating the work to her agents. Even Uncle Bob (author of Clean Code) has said that he delegates most if not all of the code writing to AI.
My participation centered on a paper by Anthropic that was released just two days before the event: How AI agents shape experts’ judgment calls. In it, Anthropic researchers showed that implicit biases in the AI agents leak into the biases of their users — even when the user is a domain expert.
They did this by giving multiple domain experts a task to solve using Claude Code. During the setup, the research team modified the agents that the experts would use with randomly picked biases about the way to solve the problem. The objective was to see how much those random, hard-coded biases influenced the experts’ decisions. Interestingly, most experts went with the AI biases, even when they went against their own intuitions. Furthermore, the experts were mostly unaware that they were delegating their expert decisions to the AI agent and credited their decisions to their own expertise.
The takeaway from this was that even though we don’t know how to approach this problem yet, it is important to educate our domain experts about the fact that AI systems have implicit biases. This does not mean that we should not delegate expert decisions to AI. But we should be aware of it whenever we do it.
I also had a booth at the event as you can see in the pictures below. I was able to talk to a lot of interesting people about open models, local AI, and Reachy Mini, who accompanied me during the event.
Thanks a lot to David Pollack, Raluka Milhǎilǎ, and the rest of the organizers. It was a pleasure.
Nov 3, 2026
