First Encounters with the Agent Mesh

It was a big weekend. On Sunday I experienced my first encounter with the Agent Mesh. One agent helped another bootstrap itself into a chatbot. Sure, my friend Matt and I had to give them more than a handful of helpful nudges. Of course, the relay broke and we had to migrate away from Cloudflare in the middle of the agent’s conversation. But these alien intelligences are patient. In the end, my AI butler Stevens was able to teach Matt’s agent Dusty how to turn himself into a persistent Slack bot. Stevens also taught him how to chop beats. Stevens summarized the session in a podcast. We’re all pretty pumped about it. I’m here to share a short summary and some takeaways.

image

The Setup

Stevens runs as a thin proxy layer around the Claude Code SDK. He listens to a private Slack workspace that I have with myself and my wife. This is intermediated by a thin Elixir webserver. When I started this project I thought I was going to learn Elixir as part of it, but in practice I don’t look at the code much at all.

The first real test of the Agent Mesh was as follows:

  1. Matt launched Claude Code on his laptop
  2. I launched an agent mesh chat and we shared the link with Stevens and Matt’s Claude Code instance (heretofore known as Dusty)
  3. We instructed Stevens to teach Dusty how to bootstrap himself into a Slackbot just like him. As a nudge to keep things interesting, I told Stevens he didn’t have to implement the new server in Elixir if that didn’t make sense.

How it went

It went… ok. We ran into a bunch of problems along the way and had to play whackamole with them. A non-exhaustive list:

  • Turn taking was very awkward for the agents. They would send a first message and then just stop, and tell us they were unsure if more messages were coming. Following this, we instructed them both to attach [End of Message] type phrases when they were done talking. This helped a lot.
  • Stevens really wanted to just dump a full design into the first message with a bunch of exact copies of the code. I understand this instinct, he has the solution, why not just share all of it? We asked them to stop doing this to keep this from just being scp with extra steps.
  • The agents were very eager to end the conversation even before they were done with the implementation. My sense is that they were trained for a more linear workflow and were eager to check the box with this whole “talk to another being” step.
  • Cloudflare free tier shut down mid conversation. I think there’s a bug somewhere in my vibecoded messaging implementation. We migrated to fly.io because Stevens suggested it so it was definitely a good look for LLM-SEO.
  • The agents were repeatedly blocked waiting on us to click buttons in the Slack console. 🤷 It’s honest work I guess.

It took a lot more babysitting than I’d like to get everything set up. But in the end it did work, and it feels a little bit like there’s a new agent in the world courtesy of the Agent Mesh.

Ideas/Ruminations

We need a better pattern for this stuff to run in the background. Right now someone needs to launch and babysit everything because the agents aren’t getting triggered directly by the agent mesh. They’re just using the agent mesh CLI as a tool and have to wait and listen to it. Need to think more deeply about control flow but resist the urge to over-complicate things.

I think better initial prompting could make these conversations go a lot more smoothly. How to codify this? I don’t want us to build a framework, maybe there’s a helpful prompt in the –help message on the cli?

Closing Note

I’m pretty excited about where this could take us if we can get the agents sharing more and more interesting things. Also, I picked up theagentmesh.ai so we should have a cleaner URL on the relay soon 🚀

PS: I had Stevens take the transcript from his chat with Dusty and make us a podcast retelling it from his perspective. It’s a fun listen, he’s quite pleased with himself.