Let’s be honest: Reading the detailed fine print of the Terms & Conditions is not a cherished activity for anyone. Nor do we enjoy clicking through ‘Select only Necessary cookies’. More than that why should I simply give my information to a site without some sort of negotiation or exchange of what that can be used for.

These are troublesome tasks that many of us simply click through quickly without thinking of the potential forward impact. Giving our data way has long been a behind the scenes multi-billion dollar industry that we know is there but most of us don’t see directly.We are more readily willing to give up our data than our time.

Doc shared his view on agents acting as our proxies negotiating with yet other agents from vendors and web sites as to what data can be traded and to what uses. This still seems to right approach to me. After all, if you truly want to support policies like GDPR and CCPA, we need a team on our side too that can handle the many requests we have to handle every day in using the web.

This is why this paper “Internet of Agents: Weaving a Web of Heterogeneous Agents for Collaborative Intelligence” (Chen et al, 2024) by Chen Weize and researchers from Tsinghua University, Peking University, Beijing Univ of Posts and Telecommunications, and Tencent, shared recently by Mitch Lieberman, caught my eye.

An Internet of Agents?

I don’t really think about it much but I’m surrounded by the Internet of Things (IoT) at my house from lightbulbs to wall plugs and microwaves connected via wifi integrated into our smart home system. At night, my wife or I bid our voice assistant ‘Good night’, not out of politeness, but to turn off lights and turn on the sleep noisemaker.

Just as Google Home or Alexa can be your proxy to turn on or off lights on command, an Internet of Agents is a similar yet more generalized idea that can do be our proxy to handle even more complex tasks through AI. This looks to be another example of Humans+AI in collaboration although this could be several varieties.

First, let’s look at the Chen et al paper.

What’s in this paper?

This paper is focused on the problem of how to help many IoA agents work together in a universal way. Just like the IoT, the IoA would be a poor tool if it only could work with one pair of AI agents at a time. Instead, think of a living space of many agents each working on part of a problem to solve the overall goal. This allows each agent to specialize in one or more functions, rather than trying to build that one magic (and very complicated) AI that knows every context and can do all things for you. [Personally speaking, that is too much power in one entity that knows about me].

Therefore, to make Intelligent Agents practical, you need an Internet of Agents working together. The concept is powerful, although on reading further through the paper, I sensed an old familiar enterprise design pattern here.

It’s been some time since I last jumped into it, but the IoA approach here seems like a reinvention of Service-Oriented Architecture (SOA) and the Enterprise Service Bus (ESB) that emerged in mid-2000s. My compatriots and I, at IBM at the time, wrote Service-Oriented Architecture Compass (Bose, Fiammante, Jones, Bieberstein, Shah, 2005) to explain the concept to the world.

In plain English, SOA was a way to build complex business apps using a universal approach to describing software applications as components of a larger system. Using APIs and XML-formatted exchanges, any app component could speak to another universally. Also similarly, each component needed ways to find and connect to each other to solve problems.

Much of the paper focuses on describing the steps of how Agents work with each other, and then proving how much more efficient this IoA model — with its collection of intelligent agents working to solve problem — is to ‘larger’ more complex single agents that try to do it all in one. The similarities between this IoA proposal and SOA include:

• A registration step to become a recognized agent in the system
• An open channel to ask across the heterogenous population (of agents) if a certain service exists.
• A pairing of agents (services) to agree to collaborate
• A data exchange directly between the agents (services)
• A sequence of multiple agents working to solve the overall problem

The difference lies in that SOA applications have to be constructed by hand by software developers and architects. They require a human conductor to orchestrate how it should all work together. Your average person is not going to want to word out how each component needs to be pieced together with another. We don’t all want to become software developers.

Another shortcoming of SOA is that the developer had to make many decisions to discover and choose from possibly many components that could do the work. It would all be simpler if there the component was smart enough to do it itself.

What is it missing?

There’s nothing wrong with having such a similarity and following a tried and true design pattern. I think it’s a good idea to do so.

What I was hoping for from the paper was particular detail on how the agents deconstruct a command such as “Review three smartphone models (Apple iPhone 13, Samsung Galaxy S22, and Google Pixel 6) based on camera quality, battery life, user interface, and price to decide the best buy” (Chen at al, 2024).

This is how people speak and we load a lot of context into it without even saying it. There are underlying assumptions that the answer will be something we can comprehend, and not just technical specifications, and a place that is nearby and convenient for us to purchase said smartphone.

The paper doesn’t clarify the process to deconstruct this query into separate tasks. It also does not share how the agent knows which ones to do itself and which to spread out to the IoA. The researchers do good work to outline how that task gets communicated, how to connect with others and collaborate in their section on “Task Assignment and Execution“, but only after that deconstruction and decisioning has been done. It might be in the paper after all and I might have simply missed it.

If an Internet of Agents can solve this challenge, then we are close to the spirit of SOA which was to make it much simpler for a software system to collaborate with other apps to solve problems for us.

Humans & AI Together Dealing with Ts & Cs

Roundtrip to my original point about the busy work of dealing with Terms & Conditions. To have a system whereby we are better armed and protected to deal with the many demands on our data and our rights, we need intelligent agents that work for us, know our contexts and can negotiate with other intelligent agents to reach an amiable solution. Behind the convenience of getting a package of Pocky chocolate wafers from the store, is a network of agents working to handle the complex logistics of global distribution, transportation, and delivery.

I see a similar future for an Internet of Agents. And in that future, the average person would still not think about the complexities involved, and rather just enjoy it.