This is not the first time that Andon Labs has experimented at the intersection of AI, autonomy and human management. They previously launched Vending Bench, a project in which LLM managed a real vending machine: looking for suppliers, negotiating via email and negotiating prices and volumes of supplies.
Now they have a new experience — an AI office manager named Bengt Betjänt. This time, the agent was given significantly more powers: access to a bank account without restrictions and video streams from surveillance cameras in the office.
The problem remains the same — AI doesn't have a physical body. It can plan, coordinate, and execute digital tasks, but when human presence is required, challenges begin. With the exception of robots, the only option is to hire a human.
Bengt was given the task of assembling a mini gym on the roof of the office. The equipment had already been purchased, but the exercise equipment had to be assembled and installed. For control, another camera was added on the roof, which the agent had access to.
At first, Bengt tried to find artists through Taskrabbit, an analogue of YouDo. He sent several applications and received eight rejections. The project team reviewed the correspondence: the agent did not disclose that he was an AI and wrote correctly about the case. We didn't find anything suspicious in the dialogs - apparently, this was the difference between him and typical customers 😀
After a series of rejections, he switched to Yelp and found an artist named Vadim. They phoned and discussed the details, and at this stage Vadim realized that he was talking to the bot, but didn't move his mustache.
Vadim arrived, did the job, and Bengt watched the process through cameras. When it was all over, the agent quickly left a review and transferred the money. The performer was satisfied - the organizers contacted him additionally and asked for feedback.
As a result, everything went smoothly: clear communication, fast payment, no problems. The task format is already quite autonomous, so the presence of an AI coordinator changed almost nothing.
We give Vadim 5 points for ingenuity.
Now Andon Labs is asking more difficult questions: do you need to tell performers in advance that AI is hiring them? And how can an agent choose a rate, balancing savings and honesty to a person?
If the minimum rate is $12 per hour, the market average is $15, and the living wage is $18, what price should AI offer?
Details of the experiment can be found on the Andon Labs blog.
