My chief of staff needed a staff

In September I was building a website and I had one agent working on the structure and another on branding. At some point they started making overlapping edits to the same files and asking me what to do. I got annoyed and requested their session IDs 1 and then I gave each the other’s ID and said “you guys need to figure this out.” And they did! I had never seen two lead agents coordinating with each other like this and it was an eye opening moment.

It reminded me that I’m managing a team of very simple geniuses who don’t REALLY know what’s going on at any moment.

I think most people understand that agents don’t actually persist over time, but I’m going to explain it my way. We personify the models because we talk to them like people. But the models and the agents they spawn aren’t like people—for one thing they don’t really have a memory. Every agent that spawns is basically a blank slate that only knows what’s written down around them 2 and whatever was in the first prompt. Any “memory” they have is a collection of markdown files that were drafted by the user or an agent that came before them. So managing context and memory is critical to getting agents to work together.

And when I first started messing around with Claude Code I was spinning up agents from any directory on my laptop and leaving a trail of context, “memories”, and handoff files across folders and projects. It was amazing because I had an agent working with me on my laptop without the confines of a GUI, but… Read more ↗