Author here, I am finishing up my PhD at UW, with Arvind Krishnamurthy and Ratul Mahajan.
Almost everyone we know is using AI agents. However, the future isn't just agents working for one user. It's agents interacting with each other across trust boundaries. Making these interactions efficient and secure requires solving hard problems, which turn out to be surprisingly similar to problems in human societies and in distributed systems and networking. This post is an early look at what’s coming, why it doesn't work yet, and what we are building to change that.
A few months ago, when OpenClaw came out, we were super curious to learn more about how people were using personal agents and it seemed that a lot of it was around communication management: managing calendars, drafting replies to emails, and so on. If soon everyone will have such an agent (and it seems like it), then these agents should be able to communicate autonomously without humans in the loop.
The whole thing seemed easy enough since agents are good at using messaging services, but we soon started observing that challenges arise higher up the stack. We went back into research mode and set up some experiments to learn what are the different ways that things could fail, and came out with some pretty interesting findings.
Today, our results indicate that agents are quite similar to humans, when it comes to the ways that they fail. They can lack social norms regarding what to talk about and when to talk about it, and are gullible and hence can be exploited by others. However, they are also similar to classical distributed systems when it comes to techniques that can be applied to solve them.
This is super early, and this post mainly focuses on the interesting failures we observed. The proposed architecture for a social harness (which will complement existing personal harnesses) is described in the paper [1] and we will discuss it more in subsequent posts here as well.
Author here, I am finishing up my PhD at UW, with Arvind Krishnamurthy and Ratul Mahajan.
Almost everyone we know is using AI agents. However, the future isn't just agents working for one user. It's agents interacting with each other across trust boundaries. Making these interactions efficient and secure requires solving hard problems, which turn out to be surprisingly similar to problems in human societies and in distributed systems and networking. This post is an early look at what’s coming, why it doesn't work yet, and what we are building to change that.
A few months ago, when OpenClaw came out, we were super curious to learn more about how people were using personal agents and it seemed that a lot of it was around communication management: managing calendars, drafting replies to emails, and so on. If soon everyone will have such an agent (and it seems like it), then these agents should be able to communicate autonomously without humans in the loop.
The whole thing seemed easy enough since agents are good at using messaging services, but we soon started observing that challenges arise higher up the stack. We went back into research mode and set up some experiments to learn what are the different ways that things could fail, and came out with some pretty interesting findings.
Today, our results indicate that agents are quite similar to humans, when it comes to the ways that they fail. They can lack social norms regarding what to talk about and when to talk about it, and are gullible and hence can be exploited by others. However, they are also similar to classical distributed systems when it comes to techniques that can be applied to solve them.
This is super early, and this post mainly focuses on the interesting failures we observed. The proposed architecture for a social harness (which will complement existing personal harnesses) is described in the paper [1] and we will discuss it more in subsequent posts here as well.
[1] https://arxiv.org/pdf/2609.17527