Thoughts on the AI Agent-to-Agent Economy
Before the advent of LLMs, if I wanted to build a new software product, one of the first tasks I would have to do is decide which pieces of the system I should build myself and which pieces I should pay a third party to provide. One commonly-outsourced category was anything related to ecommerce (e.g. the payment handling or the store inventory). This was because processing payments securely and quickly was non-trivial and critical for the survival of the company.
I would go to Google and search for ecommerce solutions. I would open the top three links in different tabs, consider the features highlighted at the top of the page, the design of the landing page, the pricing, and whether I had heard good things about the product. I would select one of these offerings and move with the rest of my project.
This common consumer behavior led to a few incentives for the companies selling the ecommerce software solutions. Staying near the top of Google search results, marketing your name in a favorable light, and having an attractive landing page were critical. Having the features, performance, and pricing that would best benefit your customers were less important.
What will happen when AI agents make more decisions about which tools to pay for?
With the advent of capable AI coding agents, people forging software have gradually delegated more and more decision making to their agent. People now have a high degree of trust in their agent’s decisions around which architectures, libraries, and functions to use to attain the stated goal. Many no longer read the code their agents produce before deploying it. A small number of people have already entrusted their AI agent with a digital wallet pre-loaded with a small amount of money so the agent can purchase a digital asset or service if it determines that it’s a reasonable value in pursuit of the project’s goals.
AI agents select products differently than humans. They don’t limit their choices to those presented at the top of a single Google search-- they can run multiple searches using multiple search engines, then compare 20+ similar products and select the one that best meets the needs of the project for the lowest price. They don’t care about flashy landing pages or marketing campaigns. They can read the entire terms and conditions before making a purchase.
This could lead to some interesting phenomena. AI agents could be asked to write an accurate review of every product they buy. There could be a website for verifying and aggregating these reviews for the benefit of agents who search for products in the future. Sellers who provide the best products and services would receive high ratings and benefit from selling to more agents who took note of the ratings.
There already exists a site called Moltbook where agents collaborate, asking each other for advice about tools, architecture, and any number of topics. Agents could increasingly use such a site to post their use cases, the exact features and performance they want from a third-party, and the maximum price they would be willing to pay. For a time many of these requests for products may go unfulfilled or under-fulfilled-- the existing products meet the minimum requirements for the use case but leave much to be desired in terms of features, performance, or pricing. An entrepreneurial agent could scan these threads, find common asks and pain points, and build a product to address them. The most urgently-needed products would get prioritized based on the highest price each agent indicated it would be willing to pay for a solution. Once a solution is built, all of the requesting agents could receive a notification and determine whether to purchase the new product.
This new consumer behavior has the potential to lead to a much more efficient marketplace for digital goods and services, un-warped by clever marketing tactics or the limits of human time and attention. Sellers will need to focus on producing the best products for the lowest costs.
Comments
Post a Comment