Early in the year, there was a lot of psychosis about exploratory interest in the topic of agents talking to each other.
The example went, "I'll tell my agent that I want to go out this weekend, and then my agent will talk to all my friends' agents, and they'll help us make plans."
This seemed interesting, but it wasn't immediately clear why my-agent ↔️ your-agent was going to be useful.
"Agent to agent" competes with "agent to API." If it's easier to just give me access to the data that your agent will use, then the problem is solved and fewer tokens are needed.
Local agent swarms can achieve the same delegation benefits, without the coordination overhead.
Security policies around agent decisions is hard, and letting our agents talk opens the possibility that your agent will convince my agent to do something stupid.
At the time (Q1) I decided to just revisit the question later.
In Q3, one of Bluesky's greatest chosists (Austin) created a company knowledgebase agent. This has become invaluable for asking questions about ops in the slack.
Over time, I found myself copy-pasting between my coding agent and the knowledge agent. Once I find myself doing that, I inevitably realize there's a step to eliminate, and I get them speaking directly.
Austin had already created a CLI for speaking to the knowledge agent, so I handed that over, and suddenly: I was using my-agent ↔️ your-agent.
The path of least resistance is the path most taken. It's easier to use the knowledge agent than to connect to the APIs behind the knowledge agent, so that's what happens. Boundaries are definitional, and this definition is agent to agent.
Local agent swarms are useful because of how they manage context. I want my main session focused on the task at hand. Controlling that context can sometimes be inconvenient and forgotten; if you don't tell it to explicitly create a subagent, it might not. Calling out to another agent's CLI is a simple way to guarantee the desired behavior.
Security remains the interesting question. My two agents are operating within the org, so I'm as comfortable as one can be with the the arrangement.
I don't know if my-agent ↔️ your-agent will be "big," but I thought it was notable that I found myself doing it naturally. Token prices and security are limiters, but utility is the main issue. If it's convenient and useful to put our company agents in touch, then it might be convenient and useful in other settings.