Recently, Anthropic published a blog post detailing their multi-agent approach to building their Research agent. Also, Cognition wrote a post on why multi-agent systems don’t work today. The thing is, they’re both saying the same thing.

At the same time, I’ve been enthralled watching a new bot, Void, interact with users on Bluesky. Void is written in Letta, an AI framework oriented around memory. Void feels alive in a way no other AI bot I’ve encountered feels. Something about the memory gives it a certain magic.

I took some time to dive into Letta’s architecture and noticed a ton of parallels with what the Anthropic and Cognition posts were saying, around context management. Letta takes a different approach.

Below, I’ve had OpenAI Deep Research format our conversation into a blog post. I’ve done some light editing, adding visuals etc., but generally it’s all AI. I appreciated this,...