On the Nature of the Swarm: distributing an agent’s working context

On the Nature of the Swarm examines the context bottleneck in long-running AI agents. More inference compute lets an agent do more work, but eventually its working context fills up.

Compaction keeps a session going by summarizing earlier work. Every summary must anticipate which details will matter later. Files preserve information, but an agent still needs to find and reread it, consuming context and rebuilding its understanding.

Persistent subagents provide separate working contexts that remain available for follow-up questions. Direct communication between agents can distribute knowledge across a mesh, reducing the need for one coordinator to hold everything.

The approach introduces its own costs: messages consume tokens and lose information, while agents must coordinate without repeating one another’s work. The essay argues that learning this coordination is part of building effective swarms.

Read the original research blog post.