AI agents may reduce transaction costs, increase congestion and disrupt markets

AI agents are promised to be the most helpful assistants that perform actions on the web on your behalf, such as booking trains, finding the best accommodation or searching for that rare and precious gift. Agents can do this fast, making the transaction costs of searching, negotiating and other economic activities negligible. Everyone who can afford such an agent will be incentivised to deploy them, causing much higher volumes of transaction, and concentrating the economy in places that generate the most value. The moment agents become clever enough, the web’s functions will forever change, with impacts far beyond some AI slop on your favorite social media platform.

The Coasean Singularity describes a future with radical changes in demand, supply and market design with AI agents. Agents don’t tire, don’t forget, and can request and analyse hundreds of quotes for products and services in seconds at near-zero marginal cost. Where it would first be impossible for you to search the web for the gems you are looking for, your agent will do it for you soon.

But this new efficiency comes with a new form of dependence. The key question is no longer whether agents can act for us, it’s who controls them. Will consumers have flexible, transparent agents that genuinely represent their preferences, or will platforms and large companies own and gatekeep the best-performing systems? Maybe like now, a few large platforms will lead, and you will have to pay whichever subscription fee that Google or Anthropic desires from you.

Many people will ask their agents to act on their behalf, leading to agent congestion. When applying for a job is automated, you can apply to as many jobs as you please, your agent will sort out the most promising positive responses anyway. The market will have to introduce new friction, such as a small cost for applying, because costless actions will lose their meaning. As a human it will become key to prove you are a person, otherwise you can just send multiple bots representing you in different ways. What will this trade-off between anonymity and verification look like? It will become critical for consumers to specify preferences, and check for misinterpretations and hallucinations. How many mistakes have we already seen with fake citations and published chatbot outputs?

Agents will likely drive forces we are not ready for, and maybe cannot even do much about.