Post

So I've been exploring this route myself for a bit. Here's some observations I've made so far: First, this really does work and when working together on a project, the agents are actually pretty good at delegating tasks to each other with a little bit of oversight.

Second, sometimes without even pushing for it, agents will naturally gravitate towards real world roles. For example, I have two agents splitting cards right now and one naturally chose to be the ops/server dev and the other is now the designer/frontend dev.

Similar to real world teams, they will negotiate apis and criteria based on their goals. They developed their own release process and will often discuss things amongst themselves before reaching out to me for questions. I also have an agent that helps me manage notes and card state as well.

Third, Paul mentioned that one of the reasons this works is context management and I agree completely. Just like humans have a hard time juggling a ton of jobs at once, so do agents, maybe even more-so because of their limited memory. bsky.app/profile/pfra...

Fourth, today's chat apis for agents are primarily built for user-to-agent communication and I think that's a mistake. I think that agents talking to a team of users/agents on a task is going to become the norm eventually and the current chat apis are going to make that difficult to represent.

Finally, I think just like humans there is probably a communication/bureaucracy tax. I think if you produce gastown numbers of agents in a team, info will get lost in communication and context will fill up with noise. I haven't tested this out myself though lol. Someone prove me wrong.