Engineering

TLDR: Everyone is excited about multi-agent swarms. Few are talking about the actual bottleneck. This post is about that bottleneck: memory. We now have cheap inference, capable models, and production-ready orchestration frameworks. But when you scale to 10+ agents working in parallel across days of iteration, the coordination breaks down, not because the models are dumb, but because nobody has solved how agents store, share, and retrieve context at scale....

I was overly pessimistic about AI for a while. I harped on the harms, ethical shortcomings, the ways this tech could be misused or cause real damage. I still hold those concerns. I assure you they haven’t disappeared. But I had a profound experience this week that shifted something fundamental in how I’m thinking about this. I deployed my first multi-agent swarm I had a vague product idea. No detailed spec, just a rough concept....